WO2010023808A1 - 画像同一性尺度算出システム - Google Patents
画像同一性尺度算出システム Download PDFInfo
- Publication number
- WO2010023808A1 WO2010023808A1 PCT/JP2009/003283 JP2009003283W WO2010023808A1 WO 2010023808 A1 WO2010023808 A1 WO 2010023808A1 JP 2009003283 W JP2009003283 W JP 2009003283W WO 2010023808 A1 WO2010023808 A1 WO 2010023808A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- quantization
- quantization index
- index
- image
- comparison
- Prior art date
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/50—Information retrieval; Database structures therefor; File system structures therefor of still image data
- G06F16/58—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually
- G06F16/583—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually using metadata automatically derived from the content
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/50—Information retrieval; Database structures therefor; File system structures therefor of still image data
- G06F16/58—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually
- G06F16/583—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually using metadata automatically derived from the content
- G06F16/5838—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually using metadata automatically derived from the content using colour
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/22—Matching criteria, e.g. proximity measures
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/20—Image preprocessing
- G06V10/28—Quantising the image, e.g. histogram thresholding for discrimination between background and foreground patterns
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/74—Image or video pattern matching; Proximity measures in feature spaces
- G06V10/761—Proximity, similarity or dissimilarity measures
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
- H04N19/00—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals
- H04N19/10—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using adaptive coding
- H04N19/102—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using adaptive coding characterised by the element, parameter or selection affected or controlled by the adaptive coding
- H04N19/124—Quantisation
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
- H04N19/00—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals
- H04N19/10—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using adaptive coding
- H04N19/102—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using adaptive coding characterised by the element, parameter or selection affected or controlled by the adaptive coding
- H04N19/124—Quantisation
- H04N19/126—Details of normalisation or weighting functions, e.g. normalisation matrices or variable uniform quantisers
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
- H04N19/00—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals
- H04N19/10—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using adaptive coding
- H04N19/134—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using adaptive coding characterised by the element, parameter or criterion affecting or controlling the adaptive coding
- H04N19/136—Incoming video signal characteristics or properties
- H04N19/137—Motion inside a coding unit, e.g. average field, frame or block difference
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
- H04N19/00—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals
- H04N19/10—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using adaptive coding
- H04N19/134—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using adaptive coding characterised by the element, parameter or criterion affecting or controlling the adaptive coding
- H04N19/136—Incoming video signal characteristics or properties
- H04N19/14—Coding unit complexity, e.g. amount of activity or edge presence estimation
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
- H04N19/00—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals
- H04N19/10—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using adaptive coding
- H04N19/169—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using adaptive coding characterised by the coding unit, i.e. the structural portion or semantic portion of the video signal being the object or the subject of the adaptive coding
- H04N19/17—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using adaptive coding characterised by the coding unit, i.e. the structural portion or semantic portion of the video signal being the object or the subject of the adaptive coding the unit being an image region, e.g. an object
Definitions
- the present invention relates to an image identity scale calculation system for calculating an identity scale indicating the degree of identity of an image, and in particular, an image identity scale calculation capable of determining an image duplicated by various modification processes as the same image as an original image. About the system.
- An image identity scale calculation system that calculates an identity scale indicating the degree to which two images are the same is used to detect, for example, a copy of an image or a moving image that is a collection of images. That is, the image identity measure calculated from the two images is compared with a certain threshold value to determine whether the two images are the same (replicated).
- modification processing When an image is copied, various modification processes are often performed. Various types of modification processing include image compression format conversion, image compression rate conversion, image size / aspect ratio conversion, image color tone adjustment, various image filter processing (sharpening, smoothing), etc. is there. In addition, local processing on an image such as superimposition of a telop is included.
- duplication not only a simple image copy but also a case where the above-described various modification processes are performed together will be simply referred to as duplication.
- various modification processes will refer to various modification processes for images as described above. For this reason, in order to detect image duplication, the identity scale calculation system is robust to various modification processes, and calculates an identity scale that can determine that an image copied by various modification processes is the same image as the original image. Is required.
- Non-patent literature 1 to non-patent literature 3 describe a method for calculating an identity scale robustly for various modification processes.
- feature quantities of two images are extracted, and feature quantities of the two images are compared to calculate an identity scale.
- a feature amount is extracted for each of a plurality of local regions of an image
- a quantization index is calculated by quantizing the extracted feature amount
- a quantization index for each local region is extracted as a feature amount of the image.
- the quantization indexes for the corresponding local regions are compared, and the identity measure is calculated based on the number of local regions whose quantization indexes match between the images.
- Non-Patent Document 1 a quantization index is obtained by classifying luminance distribution patterns in a local region into 11 types.
- Non-Patent Document 2 the color information of the local area is normalized in the time interval, and linear scalar quantization is used as the quantization index.
- the identity measure is calculated as a Hamming distance.
- Non-Patent Document 3 (a technique described as “Local Edge Representation” in Non-Patent Document 3), a quantized index is obtained by quantizing the centroid position of the edge point extracted from the local region.
- FIGS. 1 and 2 are block diagrams showing the configuration of the image identity scale calculation system described in Non-Patent Document 1 to Non-Patent Document 3.
- the image feature quantity extraction device 11 includes a feature quantity extraction unit 111 and a quantization index calculation unit 112.
- the feature amount extraction unit 111 extracts the feature amount for each of a plurality of local regions of the image specified in advance, and supplies the extracted feature amount for each local region to the quantization index calculation unit 112. To do.
- the quantization index calculation unit 112 quantizes the feature amount for each local region supplied from the feature amount extraction unit 111 to calculate a quantization index, and outputs the quantization index for each local region as an image feature amount.
- the image feature amount comparison device 12 includes a quantization index comparison unit 121.
- the quantization index comparison means 121 receives the quantization index for each local region of each of the two images output from the image feature quantity extraction device 11, compares the quantization index for each corresponding local region, and compares the quantization index. An identity measure is calculated based on the number of local regions that match and is output.
- the image identity scale calculation system shown in the configuration of FIG. 1 and FIG. 2 uses a quantization index calculated by quantizing a feature amount as a feature amount, and therefore some image signals resulting from various alteration processes to the image are generated. Robust against change. Also, since the quantization index for each local area is used as a feature quantity and the identity scale is calculated based on the number of local areas with the same quantization index, modification such as local processing on the image, such as telop overlay Also robust to processing.
- the image identity scale calculation system shown in the configuration of FIGS. 1 and 2 calculates the identity scale using a single quantization index set quantized by a single predefined quantization method. Therefore, there is a problem that the image identity determination capability is fixed by the quantization method used.
- the image identity determination ability is determined by two scales: an identification ability that is a degree at which different images can be distinguished, and a robustness that is a degree at which the quantization index does not change by various modification processes on the image, Discrimination ability and robustness are in a trade-off relationship.
- the ability to determine the identity of images is fixed, which means that the balance between discrimination ability and robustness is also fixed.
- the present invention has been invented in view of the above-mentioned problems, and the object thereof is image identity that can adjust the balance between discrimination ability and robustness, which is a measure of image identity judgment ability. It is to provide a sex scale calculation system.
- a first image identity scale calculation system is a system for calculating an identity scale indicating a degree to which two images are identical, and is hierarchically quantized for each quantization target area of the two images. Based on the information given separately, a set of quantization indexes to be used for comparison is obtained based on the separately given information. Image feature amount comparison that is selected as a comparative quantization index set, compares the hierarchical quantization index codes of the two images using the comparative quantization index set, and calculates an identity measure of the two images Means.
- a first image identity scale calculation method is a method for calculating an identity scale indicating the degree to which two images are the same, wherein the image feature quantity comparison means has a predetermined hierarchical quantization.
- a hierarchical quantization index code which is an encoding format capable of uniquely identifying the quantization indexes of a plurality of hierarchies calculated by performing hierarchical quantization for each quantization target region of the two images according to the method, is input, Based on separately provided information, a set of quantization indexes used for comparison is selected as a comparison quantization index set, and the hierarchical quantization index codes of the two images are compared using the comparison quantization index set. Then, an identity measure between the two images is calculated.
- a computer that calculates an identity measure indicating a degree of similarity between two images is calculated for each quantization target region of the two images according to a predetermined hierarchical quantization method.
- Quantization used for comparison based on information given separately, using as input the hierarchical quantization index code, which is a coding format that can uniquely identify the quantization index of multiple hierarchies calculated by hierarchical quantization
- Select a set of indexes as a comparative quantization index set compare the hierarchical quantization index codes of the two images using the comparative quantization index set, and calculate an identity measure of the two images It functions as an image feature amount comparison means.
- an image identity calculation system capable of adjusting the balance between discrimination ability and robustness, which is a measure of image identity determination ability.
- FIG. 3 is a block diagram showing the configuration of the image identity scale calculation system according to the first embodiment.
- the first embodiment of the present invention includes an image feature quantity extraction device 01 and an image feature quantity comparison device 02.
- FIG. 4 is a block diagram showing a specific configuration of the image feature quantity extraction device 01 in the image identity scale calculation system according to the first embodiment
- FIG. 5 shows the first embodiment. It is the block diagram which showed the specific structure of the image feature-value comparison apparatus 02 in the image identity scale calculation system concerning.
- the image feature quantity extraction device 01 When an image is input, the image feature quantity extraction device 01 performs hierarchical hierarchies according to a hierarchical quantization method that is a method of performing hierarchical quantization in advance for each quantization target region of a predetermined image. Quantization is performed to calculate a quantization index of a plurality of hierarchies, to calculate a hierarchical quantization index code that is an encoding format that can uniquely identify the quantization index of each hierarchy of each quantization target region, Output as image features.
- the quantization index of each layer in each quantization target region can be uniquely specified
- the correspondence between each quantization target region and its quantization index can be uniquely specified
- the quantization target region is a region to be subjected to quantization in a hierarchical manner to calculate a quantization index.
- the quantization target region may be a plurality of local regions obtained by spatially dividing an image, for example.
- a plurality of local regions (blocks) obtained by dividing an image into rectangular local regions (blocks) (the blocks may partially overlap each other) may be used.
- a local region may be defined for a normalized image obtained by normalizing the image size so as to cope with the conversion of the image size.
- the entire image may be set as one quantization target region.
- the quantization target region does not have to be a spatial region of the image, and may be a partial region in the frequency region of the image signal, for example.
- the quantization target region may be any region as long as quantization can be performed based on the image signal.
- the hierarchical quantization method means that a quantization target region is classified into a finite number of quantization indexes by a certain quantization method based on an image signal, and samples classified into each quantization index are further separated into different quantization indexes.
- This is a method of hierarchical classification into a finite number of quantization indexes by a quantization method.
- the quantization method is a type of feature amount used for quantization and a procedure / parameter for calculating the quantization index by quantizing the feature amount.
- FIG. 6 schematically shows an example of the hierarchical quantization method.
- a number represents a quantization index
- a portion surrounded by a dotted line represents a set of quantization indexes quantized by one quantization method.
- the quantization method of the lower layer is determined by the quantization index of the upper layer.
- the quantization method of the lower layer differs depending on the quantization index of the upper layer (the type of feature quantity used for quantization and the procedure / parameter for calculating the quantization index by quantizing the feature quantity) (however, Or the same quantization method).
- the quantization index is 1
- the quantization method of the lower layer is quantization method B
- the quantization index is 2
- the quantization method of the lower layer is quantum.
- the conversion method C is different.
- the feature value used by the quantization method B and the feature value used by the quantization method C may be different.
- the type of feature value used by the quantization method of the upper layer may be different from the type of feature value used by the quantization method of the lower layer.
- the number of quantization indexes classified in the lower layer may be different depending on the quantization index of the upper layer.
- the quantization index is 1, the number of quantization indexes classified in the lower hierarchy is 2, and when the quantization index is 5, classification is performed in the lower hierarchy.
- the number of quantization indexes is 4, which is different.
- the depth (number of hierarchies) of the lower hierarchy may be different depending on the quantization index of the upper hierarchy. Referring to FIG.
- the depth (number of layers) of the lower hierarchy is 3 at the maximum (4 including the highest hierarchy), and the quantization index is 3
- the depth (number of hierarchies) of the lower hierarchy is 1 (2 if the highest hierarchy is included), which are different.
- the quantization method in the highest layer needs to be a method for classifying all samples into a finite number of quantization indexes so as to be unique and unclassified.
- the quantization method for the other layers must be a method for classifying all samples classified into the quantization index of the upper layer uniquely into a finite number of indexes so that there is no unclassification.
- the quantization method A which is the quantization method of the highest hierarchy, has quantization indexes 1, 2, 3, 4, and so that all samples are uniquely and unclassified. (One sample should not be classified into quantization indexes 1 and 2 at the same time, for example).
- the quantization method G is configured so that all the samples classified into the quantization index 5 which is the quantization index of the higher hierarchy are uniquely and unclassified so that the quantization indexes 18, 19, 20 or 21 (one sample should not be classified into the quantization indexes 18 and 19 at the same time, for example).
- a plurality of different quantization methods may exist in parallel with respect to the quantization index of the upper layer (the highest level).
- a plurality of hierarchical quantization methods may exist in parallel). Referring to FIG. 6 as an example, with respect to the quantization index 4, a quantization method E and a quantization method F, which are different quantization methods, exist in parallel.
- any feature amount may be used, and the quantization index may be used with any procedure / parameter using the feature amount. May be calculated.
- a method may be used in which a scalar feature amount (for example, an average luminance value of an image) is extracted from the quantization target region and linear scalar quantization or nonlinear scalar quantization is performed on the feature amount.
- a method of extracting a feature quantity of a vector quantity for example, a luminance histogram of an image
- Parameters for scalar quantization or vector quantization may be arbitrary.
- a quantization method based on rules instead of simple arithmetic operations may be used. For example, a feature quantity related to the intensity and direction of the gradient of the image is extracted from the quantization target area, and if the intensity is less than a specified quantity, it is classified into a quantization index indicating “no gradient”, otherwise A method may be used in which the gradient direction is quantized into four directions and classified into five quantization indexes as a whole.
- the quantization method A is preferably a more robust quantization method than the quantization methods B, C, D, E, F, and G.
- the quantization method using the gradient of the luminance component of the image as the feature value is more robust than the quantization method using the color component of the image as the feature value.
- the robustness of a certain quantization method is, for example, creating a duplicate image by applying various modification processes to the image group for learning, and quantizing each quantization target region from the original image and the duplicate image by the quantization method.
- An index can be calculated and measured at a rate at which the quantization index matches in the corresponding quantization target area.
- the quantization method having a higher matching ratio is a more robust quantization method with respect to various modification processes on an image.
- the quantization method with high robustness should be placed higher than the quantization method with low robustness. It may be configured.
- the quantization method in each layer of the hierarchical quantization method is classified as equally as possible for a plurality of quantization indexes classified by the quantization method in a general image that does not assume a specific population. It is desirable (but not necessarily) that the quantization method be such. Referring to FIG. 6 as an example, when quantization is performed by the quantization method A in a general image that does not assume a specific population, each of the quantization indexes 1, 2, 3, 4, and 5 can be as much as possible. It is desirable to classify with an equal degree.
- the parameters of the quantization method may be set so that the image group for learning is classified as evenly as possible.
- the hierarchical quantization method is such that a set of quantization indexes calculated by the quantization method of the highest layer can identify sufficiently different images in a general image group that does not assume a specific population. It is desirable (but not necessarily) that the method has some discriminating ability. Referring to FIG. 6 as an example, in a general image that does not assume a specific population, a method in which the quantization indexes 1, 2, 3, 4, and 5 calculated by the quantization method A have sufficient discrimination capability. It is desirable that Here, for example, if the number of quantization indexes calculated by the quantization method of the highest layer is too small (for example, two), it is not desirable because sufficient identification capability cannot be obtained (however, this is not the case). May be).
- the number of quantization indexes calculated by the quantization method of layers other than the highest layer is smaller (that is, the degree of freedom of adjustment). Is preferred) (but not necessarily so).
- the number (5) of quantization indexes calculated by the quantization method of the highest layer is the maximum, and the number of quantization indexes calculated by the quantization method of the lower layer is that. Less (2, 3, 4) is desirable (but need not be).
- the number of quantization indexes calculated by the quantization method of the highest layer is about 10, and the quantization index calculated by the quantization method of the lower layer is 2 (that is, divided into two) ) Is a desirable example.
- the image feature quantity extraction device 01 includes a next hierarchy quantization method selection means 011, a feature quantity extraction means 012, a quantization index calculation means 013, a hierarchical quantization index code output means 014, It is composed of
- the next hierarchy quantization method selection unit 011 corresponds to the quantization index.
- the quantization method of the lower layer is selected and the quantization index is not supplied, the quantization method of the highest layer is selected, and the feature amount extraction unit 012 and the quantization index calculation unit 013 are selected for the information of the selected quantization method To supply.
- the quantization index input to the next hierarchy quantization method selection unit 011 is information fed back from the quantization index calculation unit 013 described later. Referring to FIG. 6 as an example, when a quantization index is not supplied, a quantization method A that is a quantization method of the highest layer is selected.
- the quantization method H that is a quantization method in the lower hierarchy of the quantization index 6 is selected.
- the respective quantization methods are selected and the respective quantization methods are selected. Provide information on the conversion method. Referring to FIG. 6 as an example, when the quantization index 4 is supplied, each of the quantization method E and the quantization method F existing in parallel in the lower hierarchy of the quantization index 4 is selected and supplied.
- the feature amount extraction unit 012 extracts, for each quantization target region, the feature amount used by the quantization method from the input image according to the quantization method information supplied from the next-layer quantization method selection unit 011.
- the extracted feature amount is supplied to the quantization index calculating unit 013.
- the feature quantity extraction unit 012 is a hierarchical quantization method that includes a plurality of quantization methods when there are a plurality of different quantization methods in parallel in the lower hierarchy (or the highest hierarchy) of a given quantization index.
- feature quantities used by the respective quantization methods are extracted and supplied.
- the feature quantity extraction unit 012 extracts all necessary feature quantities in advance, and outputs the feature quantity when the quantization method information is supplied from the next layer quantization method selection unit 011. It may be.
- the feature amount may be held and output as it is without being extracted again.
- the quantization index calculation unit 013 quantizes the feature amount supplied from the feature amount extraction unit 012 according to the quantization method information supplied from the next-layer quantization method selection unit 011 for each quantization target region. Calculate the quantization index.
- the calculated quantization index is not the lowest hierarchy, that is, when there is a lower quantization method, the calculated quantization index is supplied (feedback) to the next hierarchy quantization method selection unit 011. If the calculated quantization index is in the lowest hierarchy, the calculated quantization index is supplied to the hierarchical quantization index code output unit 014.
- the quantization index calculation unit 013 is not limited to the case where the quantization index of the lowest layer is calculated, and the quantization index is output to the hierarchical quantization index every time the quantization index of each layer is calculated.
- the quantization indexes 8, 9, 11, 12, 14, 15, 16, 17, 20, 21, 22, 23, 25, 26, 27, 28, 29, 30, 32, 33 , 34, 35, 36, and 37 are quantization indexes of the lowest layer, and other quantization indexes are not quantization indexes of the lowest layer.
- the quantization index calculating unit 013 has a plurality of quantization methods in the hierarchical quantization method, because a plurality of different quantization methods exist in parallel in the lower hierarchy (or the highest hierarchy) of a certain quantization index.
- quantization is performed by each quantization method, and a quantization index is calculated and supplied.
- FIG. 6 as an example, when information on the quantization method E and the quantization method F is supplied, a quantization index is calculated for each quantization method.
- the hierarchical quantization index code output unit 014 When the quantization index of the lowest layer is supplied from the quantization index calculation unit 013 for each quantization target region, the hierarchical quantization index code output unit 014, the quantization index of each layer in each quantization target region A hierarchical quantization index code, which is a coding format that can uniquely specify (each quantization index calculated by the quantization index calculating unit 013), is calculated and output.
- the hierarchical quantization index code the quantization index of each layer can be uniquely specified for each quantization target region. Any encoding format that can uniquely decode this may be used. Referring to FIG.
- the quantization index 27 when the quantization index 27 is supplied as the quantization index of the lowest hierarchy, for example, the quantization index of each hierarchy becomes the quantization indexes 2, 10, 27 in order from the highest hierarchy, Any encoding format can be used as long as the quantization index of the hierarchy can be specified as the quantization indexes 2, 10, and 27.
- the hierarchical quantization method a plurality of different quantization methods exist in parallel in the lower hierarchy (or the highest hierarchy) of a certain quantization index, so that a plurality of different lowest quantization indexes are supplied.
- any encoding format that can uniquely identify the quantization index of each layer for each layer may be used. Referring to FIG.
- quantization indexes 28 and 17 are supplied as quantization indexes of the lowest layer, quantization indexes 4, 13, and 28 that are quantization indexes of the respective layers, respectively.
- Any encoding format may be used as long as each of the quantization indexes 4 and 17 (the quantization index 4 is common) (each hierarchical quantization path) can be uniquely identified.
- the hierarchical quantization index code may be obtained by encoding only the quantization index of the lowest layer for each quantization target area, for example. This is because the hierarchical quantization method as shown in FIG. 6 is shared between the extraction side (image feature amount extraction device) and the comparison side (image feature amount comparison device), and if each can be referred to, This is because the quantization index higher than that can be uniquely identified from the quantization index of the lowest hierarchy. Referring to FIG. 6 as an example, if the quantization index of the lowest layer is, for example, the quantization index 27, only the quantization index 27 may be encoded (in accordance with the hierarchical quantization method, the quantization index 27 , Higher-order quantization indexes 2 and 10 can be uniquely identified).
- the hierarchical quantization index code may be obtained by encoding the quantization index of all layers including the quantization index of the lowest layer.
- the quantization index of the lowest layer is, for example, the quantization index 27, the quantization indexes 2, 10, and 27 of all layers may be encoded. Since the code size of the former is smaller, the former is generally preferable. In the latter case, the code size increases, but it is not necessary to reproduce the quantization index of the upper layer corresponding to the quantization index of the lower layer from the quantization index of the lowest layer when comparing the hierarchical quantization index codes. There is an advantage that the amount can be reduced.
- the hierarchical quantization index code may be any encoding format that can uniquely identify the correspondence between each quantization target region and the quantization index. For example, it may be encoded according to a predetermined order of quantization target areas. Further, for example, it may be encoded by adding information that can specify the quantization target region.
- the image feature quantity comparison device 02 receives a hierarchical quantization index code of two images calculated by the image feature quantity extraction device 01, and a population indicating characteristics of the population to which both or at least one of the two images belong Based on the characteristic information, a set of quantization indexes used for comparison is selected as a comparison quantization index set, the hierarchical quantization index codes of the two images are compared using the comparison quantization index set, and two An identity scale indicating the degree to which the images are identical is calculated and output.
- the population may be any population as long as both or at least one of the two images to which the hierarchical quantization index code is supplied as an input belong.
- the population may be a target to which the image identity scale calculation system of the present invention is applied. Thereby, the effect of the image identity scale calculation system of the present invention can be obtained for the image group of the population.
- the target is the image / moving image sharing service. Therefore, the population may be a database (a group of images included) in the shared service.
- the population may be, for example, an image group included in a classification to which both or at least one of two images belong.
- the classification is a collection of images having the same characteristics.
- a classification based on the content of the image for example, a natural image, an artifact image, a landscape image, a person image, a flower image, an animal image, etc.
- Etc a classification based on the content of the image
- the classification may be, for example, a genre (for example, news, sports, variety, drama, documentary, etc.), a broadcasting station, a broadcasting time zone, a broadcasting program, or the like.
- the population may be, for example, a database to which both or at least one of two images belong or an image group included in a specific subset thereof.
- an image group included in the image / moving image database of the service or a specific subset thereof is used as a population. Also good.
- the image identity scale calculation system of the present invention functions as a system that detects duplication of the same image in the database of the service when both of the two images belong to the database (population), for example. For example, if one of the two images belongs to the database (population) and the other is given from the outside, it functions as a query search for the same image in the database of the service.
- both or at least one of the two images is an image on the Internet
- an image group on the Internet or a subset thereof may be used as a population.
- a group of images posted to a specific site on the Internet may be used as a population.
- an image group or a subset thereof owned by the specific individual may be used as a population.
- both or at least one of the two images is a frame of a moving image
- a frame group included in a moving image or a partial section of the moving image to which both or at least one of the two images belongs may be used as a population.
- the moving image A (or a partial section of the moving image A) and the moving image B (or a partial section of the moving image B) or at least One group of frames may be used as a population.
- the population to which both or at least one of the two images to which the hierarchical quantization index code is supplied as an input does not necessarily need to be an image aggregate in which the images are included as one image. Any image aggregate having the same characteristics may be used. For example, if the type based on the content of an image to which a hierarchical quantization index code is supplied as an input is “natural image”, and the image group of “natural image” is a population, The image group of “images” does not necessarily include an image to which a hierarchical quantization index code is supplied as an input.
- the entire database including the image may be included in the population, but for example, a portion of the database not including the image
- the set may be a population.
- the population characteristic information is information indicating the characteristics of the population, and in particular, based on an identity scale calculated using a set of selected quantization indexes (as candidates) in the image group of the population. This information is correlated with the accuracy of the determination of the identity of the images (determination of whether or not the two images are the same). Based on this information, the comparative quantization index set can be selected so that the accuracy of the image identity determination based on the calculated identity scale is high. Note that the accuracy of image identity determination based on the calculated identity scale can be considered as the reliability of the identity scale. For this reason, the population image information may be considered as information having a correlation with the reliability of the identity scale calculated using the set of selected quantization indexes in the image of the population.
- the accuracy of the determination of the identity of the image based on the identity measure calculated using the set of selected quantization indexes is (1) it is possible to identify different images included in the set of selected quantization indexes. It is determined by two scales, that is, the discriminating ability that is the degree, and (2) robustness that is the degree that the quantization index is not changed by various modification processes to the image that the set of the selected quantization index has.
- the higher the discrimination ability and the higher the robustness the higher the accuracy of image identity determination based on the calculated identity scale (note that there is a trade-off between discrimination ability and robustness).
- the population characteristic information may be information correlated with the discrimination ability of the set of quantization indexes, which is the degree to which different images can be discriminated, in the image group of the population. Based on this information, it is possible to select a comparative quantization index set so as to increase the discrimination capability.
- the population characteristic information is information that correlates with robustness, which is the degree to which the quantization index does not change due to various modifications to the image included in the selected set of quantization indexes in the image group of the population. May be. Based on this information, a comparative quantization index set can be selected so that robustness is high.
- the information correlated with the discrimination ability of the set of quantization indexes selected in the image group of the population may be, for example, the appearance frequency for each quantization index in the image group of the population (in this case)
- the frequency of appearance for each quantization index in the image group of the population is the population characteristic information). This is because, in general, when the number of quantization indexes included in the set of quantization indexes is the same, the identification frequency becomes higher as the appearance frequency for each quantization index is equal. Based on this information, it is possible to select a comparative quantization index set so that the appearance frequencies for each quantization index are evenly approached, so that the discrimination ability is increased.
- the population characteristic information is not the frequency of appearance for each quantization index in the image group of the population, for example, information indicating a bias in the appearance frequency to a specific quantization index in the image group of the population, or indirect It may be information that can be estimated. This is because if the selected set of quantization indexes has a bias in the frequency of appearance in a specific quantization index (the frequency of occurrence of a specific quantization index is high), the identification ability is low due to that quantization index. (This is because a quantization index having a high appearance frequency has a higher probability of appearing in common among different images by chance, making it difficult to identify different images).
- the population characteristic information may be, for example, information indicating that “the image group of the population has many images including“ sky ””.
- the appearance frequency it can be estimated that there is a bias in the appearance frequency to a specific quantization index (in this case, specifically, a quantization index indicating that the color is blue, for example). This is because the identification ability is lowered due to the index.
- a specific quantization index in this case, specifically, a quantization index indicating that the color is blue, for example.
- the population characteristic information is an appearance frequency for each quantization index in the image group of the population will be described in detail in the second embodiment.
- Information that correlates with the robustness of the set of quantization indexes selected in the image group of the population includes, for example, quantization by various modification processes to the image for each quantization index in the image group of the population It may be a modification invariance indicating the degree that the index does not change (in this case, the modification invariance for each quantization index in the image group of this population becomes population characteristic information). This is because the higher the modification invariance for each quantization index, the higher the robustness. Based on this information, it is possible to select a comparative quantization index set so that the modification invariance with respect to the entire set of quantization indexes is high, and to improve robustness. In addition, the population characteristic information is not the modification invariance for each quantization index in the image group of the population.
- the modification invariance of a specific quantization index in the image group of the population Information may be used.
- the population characteristic information may be, for example, information indicating that “the image group of the population has many grayscale images”. Based on this information, it can be estimated that the modification invariance of a specific quantization index (in this case, specifically, a quantization index calculated by a quantization method using color information, for example) is low. This is because the robustness is lowered due to the index.
- the population characteristic information is the modification invariance for each quantization index in the image group of the population will be described in detail in the fourth embodiment.
- the population characteristic information is, for example, when an image is converted when the image is added to the population (for example, image compression, image size conversion).
- information obtained by learning using the images may be used.
- the population is a posting site on the Internet, information that shows the type and degree of modification obtained by comparing the image before posting with the image after posting and converting the image, and quantization It may be the modification invariance for each index.
- the image feature quantity comparison device 02 includes a comparison quantization index set selection unit 021, a comparison quantization index acquisition unit 022, and a quantization index comparison unit 023.
- the comparison quantization index set selection unit 021 selects a set of quantization indexes used for comparison as a comparison quantization index set based on the input population characteristic information. Information on the selected comparison quantization index set is supplied to the comparison quantization index acquisition unit 022.
- the selected comparison quantization index set needs to be a set of quantization indexes such that each quantization target region is uniquely classified into a quantization index. That is, when a quantization index with a hierarchical quantization method is selected, a higher-level quantization index (an ancestor node) corresponding to the quantization index or a lower-layer quantization index ( Do not select descendant nodes.
- a higher-level quantization index an ancestor node
- a lower-layer quantization index Do not select descendant nodes.
- the quantum calculated by the parallel quantization method exists. Do not select the index. Referring to FIG.
- the comparative quantization index set is ⁇ 1, 2, 3, 4, 5 ⁇ or ⁇ 6, 7, 2, 3, 16, 17, 30, 31, 19, 20 , 21 ⁇ , ⁇ 22, 23, 7, 8, 9, 26, 27, 3, 16, 17, 30, 36, 37, 19, 20, 21 ⁇ , ⁇ 22, 23, 34, 35, 8 , 9, 26, 27, 11, 12, 28, 29, 14, 15, 30, 36, 37, 32, 33, 20, 21 ⁇ , etc. It is.
- the quantization index 24 and the quantization index 7 corresponding to the higher-level quantization index corresponding to the quantization index 24 can be selected at the same time, or the quantization index 24 corresponding to the quantization index corresponding to the two-level higher-order hierarchy 1 cannot be selected at the same time.
- the quantization index 7 and the quantization index 24 or 25 corresponding to the lower layer corresponding to the quantization index 7 can be selected at the same time, or the quantization index 7 and the quantization index corresponding to the lower hierarchy of the two layers can be selected.
- the quantization indexes 34 and 35 cannot be selected at the same time.
- the selected comparison quantization index set is preferably a set of quantization indexes such that each quantization target region is classified into one of the quantization indexes so that there is no unclassification, This is not necessarily so. Referring to FIG.
- the comparative quantization index set is ⁇ 1, 2, 3, 4, 5 ⁇ or ⁇ 6, 7, 2, 3, 16, 17, 30, 31, 19, 20 , 21 ⁇ , ⁇ 22, 23, 7, 8, 9, 26, 27, 3, 16, 17, 30, 36, 37, 19, 20, 21 ⁇ , ⁇ 22, 23, 34, 35, 8 , 9, 26, 27, 11, 12, 28, 29, 14, 15, 30, 36, 37, 32, 33, 20, 21 ⁇ , etc. It is desirable because it is classified.
- the comparison quantization index set is selected as ⁇ 1, 2, 3, 4 ⁇ , the sample classified into the quantization index 5 and the quantization index of the lower hierarchy corresponding thereto becomes unclassified, preferably There is no such choice.
- the comparison quantization index set selection unit 021 selects a comparison quantization index set based on the population characteristic information. For example, when the population characteristic information is information correlated with the accuracy of the image identity determination based on the identity scale calculated using the selected set of quantization indexes in the image group of the population, Based on this information, the comparison quantization index set may be selected so that the accuracy of the determination of the identity of the image based on the calculated identity scale is increased. In addition, for example, when the population characteristic information is information that has a correlation with an identification ability that is a degree that a set of quantization indexes selected can identify different images in the image group of the population, based on this information The comparison quantization index set may be selected so that the discrimination capability is high.
- a lower-level quantization index may be selected.
- the population characteristic information is information indicating a bias in appearance frequency to a specific quantization index in the image group of the population (information indicating that the appearance frequency of a specific quantization index is high), or indirectly Is information that can be estimated (for example, information indicating that “the image group of the population has many images including“ sky ””), in order to increase the identification ability, the specific quantization index (population When the characteristic information is information indicating that “the image group of the population has many images including“ sky ””, a quantization index for a typical sky region, for example, a quantization index indicating that the color is blue
- the comparison quantization index set is selected by selecting the quantization index calculated by the lower-layer quantization method for (, becomes a specific quantization index) (A quantized index with a high appearance frequency has a higher probability of appearing in common between different images, making it difficult to distinguish different images and lowering the discrimination ability.
- the quantization index By selecting the quantization index, it is possible to divide the appearance frequency to suppress the deviation of the appearance frequency and to increase the identification ability).
- FIG. 6 for example, when a set of quantization indexes ⁇ 1, 2, 3, 4, 5 ⁇ of the highest hierarchy is first used as a candidate, an image of the population is based on the population characteristic information.
- a quantization index 11 which is a quantization index calculated by a quantization method of a lower hierarchy with respect to the quantization index 3, 12 may be selected, and the comparative quantization index set may be ⁇ 1, 2, 11, 12, 4, 5 ⁇ .
- the population characteristic information is the appearance frequency for each quantization index in the image group of the population will be described in detail in the second embodiment.
- the population characteristic information is information that correlates with robustness, which is the degree to which the quantization index does not change due to various modifications to the image, which the selected set of quantization indexes has in the image group of the population.
- a comparative quantization index set may be selected based on this information so as to increase robustness. In order to increase the robustness, for example, a higher-level quantization index may be selected.
- the population characteristic information is information that can estimate the modification invariance of a specific quantization index in the image group of the population (for example, information indicating that “the image group of the population has many grayscale images”)
- the specific quantization index if the population characteristic information is information indicating that there are many grayscale images in the image group of the population
- color information The quantization index set for comparison may be selected by selecting a higher-level quantization index for the quantization index calculated by the quantization method using By selecting a quantization index, multiple quantization indexes in the same layer are integrated, and a quantization index across the multiple quantization indexes is integrated. The change in the box is eliminated, is increased robustness). Note that the case where the population characteristic information is the modification invariance for each quantization index in the image group of the population will be described in detail in the fourth embodiment.
- the comparison quantization index set selection unit 021 has, as the population characteristic information, information correlated with the discrimination ability of the set of selected quantization indexes and information correlated with robustness in the population image group.
- the comparison quantization index set may be selected in consideration of both. Discrimination ability and robustness are in a trade-off relationship (generally, the discrimination ability increases as the discrimination ability increases, and the discrimination ability decreases as the robustness increases). For example, there is each of discrimination ability and robustness
- the comparative quantization index set may be selected so as to be close to a certain criterion, and for example, the comparison quantum is set so that the other becomes higher while one of the discrimination ability and robustness meets the certain criterion.
- a quantized index set may be selected, or a comparative quantized index set may be selected based on, for example, a scale obtained by quantifying and discriminating each of discrimination ability and robustness.
- the comparison quantization index set selection unit 021 selects a comparison quantization index set in consideration of information that does not depend on the population and correlates with discrimination ability and robustness. May be.
- information includes, for example, the number of quantization indexes included in the set of selected quantization indexes (in general, the greater the number of quantization indexes, the higher the identification ability and the lower the robustness), and the selection The depth of the set of quantized indexes. For example, from the set of quantization indexes that satisfy a certain criterion for the number of quantization indexes (for example, the number of quantization indexes is about 10 or the like), the discriminating ability and robustness are enhanced based on population characteristic information. Alternatively, a comparative quantization index set may be selected.
- comparison quantization index set selection unit 021 selects the optimum based on any of the above-described criteria determined using the population characteristic information from among the sets of quantization indexes of all combinations as candidates.
- a set of quantization indexes may be selected as a comparative quantization index set.
- the comparison quantization index set selection unit 021 selects, for example, the quantization index of the upper layer of the hierarchical quantization method in preference to the quantization index of the lower layer, so that the comparison quantization index is selected. A set may be selected.
- the comparison quantization index set selection unit 021 selects, for example, a set of quantization indexes in order from the highest hierarchy of the hierarchical quantization method, and sets a predetermined standard (for example, the above-described reference) Until a set of quantization indexes satisfying any one of the criteria) appears, a lower-level quantization index may be selected as a candidate, and a comparison quantization index may be selected.
- a lower-level quantization index may be selected as a candidate, and a comparison quantization index may be selected.
- the quantization index of the lower hierarchy based on the information indicating the bias in the appearance frequency to the specific quantization index in the population image group, the lower hierarchy of the specific quantization index is selected. You may make it select the quantization index calculated with a quantization method. For example, referring to FIG. 6, for example, first, the set ⁇ 1, 2, 3, 4, 5 ⁇ of the quantization index of the highest hierarchy is selected as a candidate, and this set is determined based on the population characteristic information.
- a lower-layer quantization index may be selected as a candidate.
- the quantization calculated by the quantization method of the lower hierarchy with respect to the quantization index 3 The quantization indexes 11 and 12 that are indexes may be selected, and the set of quantization indexes ⁇ 1, 2, 11, 12, 4, 5 ⁇ may be candidates. Such an operation may be recursively repeated until the discriminating ability determined based on the population characteristic information satisfies a prescribed standard, and the comparison quantization index set may be selected.
- the lower-layer quantization index when determining whether or not to select a lower-level quantization index, if the robustness determined based on the population characteristic information does not satisfy a prescribed criterion, the lower-layer quantization index is The comparison quantization index may be selected without selecting it.
- comparison quantization index set selection unit 021 will be described in detail in the second embodiment, the third embodiment, the fourth embodiment, and the fifth embodiment. Explained.
- the comparison quantization index set selection unit 021 does not perform processing at the timing when the hierarchical quantization index codes of the two images are input to the image feature quantity comparison device 02, but in advance the identity measure of the present invention. It is desirable to perform processing on the population to which the calculation system is applied. It is desirable that the comparison quantization index set selection unit 021 performs processing only once for a certain population (note that it is effective to perform processing periodically by updating the population) There is no need to perform processing each time a hierarchical quantization index of an image is newly input. By doing so, the image feature quantity comparison device 02 can receive a large number of hierarchical quantization index codes of a large number of images belonging to the same population (that is, for a large number of images belonging to the same population). Thus, when the identity scale is calculated), the identity scale can be calculated efficiently and at high speed.
- the comparison quantization index acquisition unit 022 receives the hierarchical quantization index code of each of the two images (image 1 and image 2) calculated by the image feature quantity extraction device 01, and for each image, For each quantization target region, information on the comparison quantization index set supplied from the comparison quantization index set selection unit 021 is selected from the quantization indexes of each layer uniquely identified by the hierarchical quantization index code. The quantization index included in the comparison quantization index set shown is acquired. The quantization index for each quantization target region acquired for each image is supplied to the quantization index comparison unit 023 as a comparison quantization index.
- the hierarchical quantization method When acquiring the quantization index included in the comparison quantization index set, if the hierarchical quantization index code is obtained by encoding only the quantization index of the lowest hierarchy, the hierarchical quantization method is used. What is necessary is just to obtain
- the quantization index of the lowest hierarchy of a certain quantization target area is 34
- the quantization index of each hierarchy is 1, 7 in the order of the upper hierarchy according to the hierarchical quantization method. , 24, 34. Since the quantization index 7 is included in the comparison quantization index, the quantization index 7 is acquired as the comparison quantization index of the quantization target region. Further, when the quantization index of the lowest hierarchy of a certain quantization target region is 27, the quantization index of each hierarchy can be obtained as 2, 10, 27 in order of the upper hierarchy according to the hierarchical quantization method. Since the quantization index 2 is included in the comparison quantization index, the quantization index 2 is acquired as the comparison quantization index of the quantization target region.
- the hierarchical quantization index code when a plurality of different quantization methods exist in parallel in the lower hierarchy (or the highest hierarchy) of a certain quantization index, the hierarchical quantization index code is one
- the quantization included in the comparison quantization index set is selected from the quantization indexes of each hierarchy of the respective hierarchical quantization paths. Get the index.
- the comparative quantization index set is ⁇ 6, 7, 2, 3, 16, 17, 30, 31, 19, 20, 21 ⁇ , for example.
- the quantization index of the lowest layer of a certain quantization target region is 29 and 17 (from the hierarchical quantization index code)
- the quantization index of each former layer is 4 in order of the upper layer.
- 13, and 29, and the quantization index of each of the latter layers can be obtained as 4, 17 in order of the upper layer (the quantization index 4 is common).
- the quantization index 17 is included in the comparison quantization index
- the quantization index 17 is acquired as the comparison quantization index of the quantization target region. Note that if the comparison quantization index set supplied from the comparison quantization index set selection unit 021 is a set in which there is unclassification, the quantum for which the comparison quantization index is unclassified (undefined). The target area will be generated.
- the comparison quantization index set is ⁇ 1, 2, 3, 4 ⁇
- the quantization index of the lowest hierarchy of a certain quantization target region is 32
- the quantization index of each layer is obtained as 5, 19, and 32 in the order of the upper layer, but none of them is included in the comparison quantization index set.
- the quantization index is unclassified (undefined).
- the comparison quantization index acquisition unit 022 calculates the identity measure for the image belonging to the population among the two images (image 1 and image 2) to which the hierarchical quantization index code is input. It is desirable to acquire the comparison quantization index in advance at the stage where the information of the comparison quantization index set is prepared, instead of acquiring the comparison quantization index immediately before performing. That is, it is desirable to obtain a comparative quantization index in advance for the image group of the population. By doing so, the image feature quantity comparison device 02 can calculate the identity scale efficiently and at high speed. This also eliminates the need to acquire a comparative quantization index every time when the identity measure is repeatedly calculated for the same image.
- a quantization index for comparison is acquired in advance for the images belonging to the database. This eliminates the need to acquire a comparative quantization index every time a query is given from the outside, and thus the identity scale can be calculated at high speed.
- a database is a population and both images belong to the database, and the identity scale is calculated for all image pairs in the database, comparison is made in advance for all images belonging to the database. By acquiring the quantization index for the image, it is not necessary to acquire the comparison quantum index every time the identity scale is calculated for each image pair, so that the speed of calculating the identity scale can be greatly increased. it can.
- the quantization index comparison unit 023 compares the comparison quantization indexes of the two images (image 1 and image 2) supplied from the comparison quantization index acquisition unit 022 for each corresponding quantization target region. Based on the number of quantization target regions having the same quantization index, an identity measure indicating the degree to which the two images are identical is calculated and output.
- the identity measure may be, for example, the number of quantization target regions having the same quantization index.
- the identity scale may be calculated based on the number of quantization target areas (Humming distances) that do not match the quantization indexes, for example.
- the identity scale may be a value obtained by, for example, obtaining the number of matching quantization target regions for each quantization index, and adding a weight value determined in advance for each quantization index.
- the identity scale is quantified by, for example, determining whether or not the quantization indexes match for each quantization target region (for example, 1 if they match, 0 if they do not match), and for each quantization target region in advance
- a predetermined weight value (for example, when the quantization target region is a local region of the image, the weight value is increased as the local region is closer to the center of the image) may be applied and added.
- the quantization is performed. The target area may not contribute to the calculation of the identity scale.
- the hierarchical quantization method, the population characteristic information, and the comparison quantization index set selected by the comparison quantization index set selection unit 021 described so far are common to all quantization target regions.
- the quantization target region or a group of a plurality of quantization target regions are common to all quantization target regions.
- the hierarchical quantization method may be different for each quantization target region or for each group of a plurality of quantization target regions.
- population characteristic information supplied as an input may be different for each quantization target region or for each group of a plurality of quantization target regions.
- the comparison quantization index set selection unit 021 may select different comparison quantization index sets for each quantization target region or for each group of a plurality of quantization target regions. In this way, for each quantization target region or for each group of a plurality of quantization target regions, the hierarchical quantization method, the population characteristic information, or the comparison quantum selected by the comparison quantization index set selection unit 021 By making the quantization index sets different, optimization suitable for the characteristics of each quantization target region or each group of a plurality of quantization target regions can be performed.
- a local region belonging to the central region of the image is a hierarchical quantization method configured by a quantization method suitable for describing a foreground object (for example, a quantization method using edge information as a feature amount).
- the local region belonging to the peripheral region of the image may use a hierarchical quantization method composed of a quantization method suitable for describing the background (for example, a quantization method using color information as a feature amount).
- FIG. 7 is a flowchart showing the operation of the image feature quantity extraction device 01 in the first embodiment.
- next hierarchy quantization method selection unit 011 selects the quantization method of the highest hierarchy according to the hierarchical quantization method for the first quantization target region, and sets the information of the selected quantization method as the feature quantity. This is supplied to the extraction unit 012 and the quantization index calculation unit 013 (step A01).
- the feature quantity extraction unit 012 determines the feature quantity used by the quantization method for the current quantization target area according to the quantization method information supplied from the next hierarchy quantization method selection unit 011. Then, the extracted feature quantity is extracted from the input image, and the extracted feature quantity is supplied to the quantization index calculating means 013 (step A02).
- the quantization index calculating unit 013 applies the feature quantity supplied from the feature quantity extracting unit 012 to the quantization index target region that is the current target. Quantization is performed according to the information of the quantization method, and a quantization index is calculated (step A03). It is determined whether or not the calculated quantization index is the lowest hierarchy of the hierarchical quantization method (step A04). If the calculated quantization index is not the lowest layer, the quantization index calculation unit 013 supplies (feeds back) the calculated quantization index to the next layer quantization method selection unit 011 to select the next layer quantization method.
- the unit 011 selects a quantization index in a lower layer of the supplied (feedback) quantization index (upper layer), and sends information on the selected quantization method to the feature amount extraction unit 012 and the quantization index calculation unit 013. Supply (step A05). Then, the process again proceeds to step A02. If it is determined in step A04 that the calculated quantization index is the lowest hierarchy, it is next determined whether or not the processing of all the quantization target regions has been completed (step A06). If all the quantization target regions have not been processed, the next quantization target region is determined, and the process proceeds to step A01 again to perform processing on the next quantization target region. If all the quantization target areas have been processed, the hierarchical quantization index code output unit 014 then encodes a quantization format that can uniquely identify the quantization index of each hierarchy in each quantization target area. Is calculated and output (step A07).
- FIG. 8 is a flowchart showing the operation of the image feature quantity comparison device 02 in the first embodiment.
- the comparison quantization index set selection unit 021 selects a comparison quantization index set based on the input population characteristic information, and compares the selected comparison quantization index set information with the comparison quantization index. It supplies to the acquisition means 022 (step B01). Note that step B01 is not processed at the timing when the hierarchical quantization index codes of the two images are input to the image feature quantity comparison device 02, but is applied to the population to which the identity scale calculation system of the present invention is applied in advance. It is better to process it.
- the comparison quantization index acquisition unit 022 calculates the comparison quantization index set supplied from the comparison quantization index set selection unit 021 from the hierarchical quantization index codes of the two input images.
- a comparison quantization index is acquired for each image, and the acquired comparison quantization index is supplied to the quantization index comparison unit 023 (step B02).
- the quantization index comparison unit 023 compares the comparison quantization indexes of the two images supplied from the comparison quantization index acquisition unit 022, calculates the identity measure, and outputs it (step B 03).
- an image identity calculation system capable of adjusting the balance between discrimination ability and robustness, which is a measure of image identity determination ability.
- the reason is that the ability to determine the identity of two images depends on which quantization index of a plurality of layers calculated for each quantization target region of the two images is used for comparison. Because it will change. Specifically, by using a lower-level quantization index for comparison, it is possible to enhance the discrimination ability, which is the degree that two images can be distinguished, out of the ability to judge the identity of two images. On the other hand, by using the higher-level quantization index for comparison, the robustness, which is the degree that the quantization index does not change due to various alteration processes to the image, of the ability to determine the identity of two images is improved. be able to.
- a plurality of hierarchies quantization index calculated based on a hierarchical quantization method is used as a feature amount, and the image feature amount comparison device 02 calculates the identity measure of a population to which at least one of two images belongs.
- the comparison quantization index set selection unit 021 selects the comparison quantization index set so that the accuracy of the determination of the identity of the image based on the identity scale calculated by the comparison quantization index set selection unit 021 is increased.
- Each has the effect of optimizing the accuracy of image identity determination based on the calculated identity measure. Thereby, the accuracy of the determination of the identity of the image based on the calculated identity scale is not greatly reduced by the image population.
- the image feature quantity comparison apparatus 02 adaptively optimizes the population and calculates the identity measure, thereby extracting the feature quantity ( That is, the hierarchical quantization method can be made common without depending on the population. For this reason, there is an effect that it is not necessary to learn an optimum feature amount extraction method (that is, hierarchical quantization method) for each population. In addition, as a result, there is no problem even when the population to which the image comparison is performed is not yet determined at the time of extracting the feature amount of the image.
- FIG. 9 is a block diagram illustrating a configuration of an image feature amount comparison apparatus of the image identity scale calculation system according to the second embodiment.
- population quantization index appearance frequency calculating means 03 is added to the configuration of the first embodiment shown in FIG. 4 and FIG.
- the difference is that the comparison quantization index set selection unit 021 is replaced with a comparison quantization index set selection unit 021A. Since the other points are the same as in the first embodiment, only the points different from the first embodiment will be described here.
- Population quantization index appearance frequency calculation means 03 receives the hierarchical quantization index code of the population image group calculated by the image feature quantity extraction device 01, and uses the hierarchical quantization method for the population image group. The appearance frequency of each quantization index is calculated, and the appearance frequency for each quantization index in the calculated population is supplied to the comparison quantization index set selection unit 021A as the population quantization index appearance frequency.
- the image group of the population may be all the images included in the population, or may be only the image group extracted by sampling appropriately as long as the characteristics of the population are reflected.
- Population quantization index appearance frequency calculation means 03 calculates all the hierarchical quantization methods from the quantization index of each hierarchy uniquely identified by each of the hierarchical quantization index codes of the input population image group.
- the appearance frequency is calculated for the quantization index. 6 as an example, the population quantization index appearance frequency calculation unit 03 calculates the appearance frequencies of all the quantization indexes 1 to 37. If the hierarchical quantization index code is obtained by encoding the quantization index of the lowest hierarchy, the quantization index of the higher hierarchy is obtained from the quantization index of the lowest hierarchy according to the hierarchical quantization method. In addition to calculating the appearance frequency. Referring to FIG. 6 as an example, when the quantization index of the lowest hierarchy of a certain quantization target area of a certain image is 34 from the hierarchical quantization index code of the image group of the population, for example, the quantization index 1, For each of 7, 24, and 34, the appearance frequency is added once.
- the appearance frequency of a certain quantization index is the sum of the appearance frequencies of all the quantization indexes in the next lower hierarchy.
- the appearance frequency of the quantization index 24 is the sum of the appearance frequencies of the quantization indexes 34 and 35
- the appearance frequency of the quantization index 7 is the sum of the quantization indexes 24 and 25
- the appearance frequency of the quantization index 1 is This is the sum of the quantization indexes 6 and 7.
- the appearance frequency for each quantization index output by the population quantization index appearance frequency calculation unit 03 may not be strictly the appearance frequency itself but may be, for example, the appearance probability.
- the appearance probability for each quantization index is a value obtained by dividing each appearance frequency for each quantization index aggregated by the above method by the number of all quantization target regions.
- the comparison quantization index set selection unit 021A uses the comparison quantization index appearance frequency calculation unit 03 so that the quantization index appearance frequencies are evenly approached based on the population quantization index appearance frequency supplied from the population quantization index appearance frequency calculation unit 03. Select a quantization index set. Information on the selected comparison quantization index set is supplied to the comparison quantization index acquisition unit 022.
- the comparison quantization index set selection unit 021 according to the first embodiment is such that the input population characteristic information is specifically the population quantization index appearance frequency
- the comparison quantization index set selection unit 021A The conditions relating to the selection of the comparison quantization index set are the same as those described in the first embodiment.
- the comparison quantization index set selection unit 021A can increase the identification ability, which is the degree to which different images can be identified, by selecting the comparison quantization index so that the appearance frequency of the quantization index approaches evenly. . This is because, when the number of quantization indexes included in the set of quantization indexes is the same, if there is a bias in a specific quantization index (when the frequency of occurrence of a specific quantization index increases), a quantum having a high appearance frequency is obtained. This is because there is a high probability that the digitized index will appear in common between different images by chance, and the discrimination ability will be low.
- the comparison quantization index set selection unit 021A selects, for example, a candidate as a method for selecting the comparison quantization index so that the appearance frequency of the quantization index approaches evenly based on the population quantization index appearance frequency.
- a set of quantized indexes in which the appearance frequency equality indicating the degree of appearance of the quantization index is calculated and the appearance frequency equality is increased based on the appearance frequency equality is calculated.
- the comparison quantization index set selection unit 021A calculates the appearance frequency equality degree for each of the combinations of the quantization index sets of all the candidates, and the quantization index that maximizes the appearance frequency equality degree. May be selected as a comparative quantization index set.
- the appearance frequency uniformity degree may be calculated based on, for example, the variance of the appearance frequency (or appearance probability) of the quantization index.
- a set of quantization index sets selected as candidates is represented as S, and the number of quantization indexes (that is, the number of elements in the set) included in the set of quantization indexes selected as candidates is represented as N.
- Is expressed as i and the appearance frequency of the quantization index i is expressed as F i
- the variance V of the appearance frequency of the quantization index can be calculated by the following equation.
- Fa is an average value of appearance frequency.
- any calculation method may be used (that is, the appearance frequency dispersion V).
- the appearance frequency equality degree may be calculated as the reciprocal of the variance V of the appearance frequency.
- the appearance frequency uniformity degree is based on, for example, the appearance frequency (or appearance probability) of the quantization index and the average value (or average value of the appearance probability) of the quantization index based on the population quantization index appearance frequency. May be calculated based on the number of quantization indexes that are equal to or less than a predetermined value (that is, the number of quantization indexes that are distant from the average value), or the ratio thereof. In this case, the greater the number of quantization indexes that are equal to or less than the prescribed value, the higher the appearance frequency uniformity degree.
- the appearance frequency uniformity degree may be any calculation method as long as the appearance frequency uniformity degree is calculated so as to increase as the number of quantization indexes equal to or less than a predetermined value increases (ie, Any method that can be calculated based on a monotonically increasing function for this number) may be used. For example, the number of quantization indexes that are equal to or less than a prescribed value may be used as the appearance frequency uniformity degree.
- the appearance frequency uniformity degree may be calculated based on information entropy (average information amount) calculated based on, for example, the population quantization index appearance frequency.
- information entropy average information amount
- the information entropy H can be calculated by the following equation.
- the appearance frequency equality may be any calculation method as long as the information entropy H is calculated to be higher as the information entropy H is larger (that is, for the information entropy H, Any method that is calculated based on a monotonically increasing function may be used.
- the information entropy H itself may be the appearance frequency uniformity degree.
- the method for calculating the appearance frequency equality is any method as long as it is a method for calculating a scale indicating the degree of the appearance frequency of the quantization index being equal. It may be a simple method.
- the comparison quantization index set selection unit 021A for example, in addition to the above-described appearance frequency equality, for example, the number of quantization indexes included in the set of quantization indexes selected as candidates (or the set of selected quantization indexes)
- the quantization index set for comparison may be selected in consideration of both the depth of the hierarchy). This is because the number of quantized indexes (or the depth of the hierarchy) is quantized as a set of quantized indexes as information independent of the equality of the appearance frequency of quantized indexes and as information independent of the population.
- the comparison quantization index set selection unit 021A in addition to the above-described appearance frequency equality, for example, for each quantization index, a modification invariance indicating a degree that the quantization index does not change by various modification processing to the image, In consideration of both, a comparative quantization index set may be selected.
- the modification invariance for each quantization index is, for example, created in advance by performing various modification processes on the learning image group to create a duplicate image, and using the image feature quantity extraction device 01 from the original image and the duplicate image. It is possible to calculate a static quantization index code and measure the ratio of the quantization index in the corresponding quantization target region by measuring for each quantization index.
- the comparison quantization index set selection unit 021A has, for example, a higher degree of appearance frequency uniformity and a modification invariance (for example, a quantization index) of the entire set of quantization indexes calculated from the modification invariance for each quantization index.
- a set of quantization indexes that increases the average value of the modification invariance for each of them may be selected as a comparison quantization index set.
- the comparison quantization index set selection unit 021A does not include, for example, the comparison quantization index so that the appearance frequency uniformity degree is high and the quantization index having a low modification invariance for each quantization index is not included.
- a set may be selected.
- the comparison quantization index set selection unit 021A selects a comparison quantization index based on the population quantization index appearance frequency so that the appearance frequency of the quantization index approaches evenly.
- the quantization index may be selected. Referring to FIG. 6 as an example, for example, when the quantization index set ⁇ 1, 2, 3, 4, 5 ⁇ of the highest hierarchy is first used as a candidate, the quantum quantization index appearance frequency is used as a candidate.
- the comparison quantization index set may be ⁇ 1, 2, 11, 12, 4, 5 ⁇ .
- the comparison quantization index set selection unit 021A also uses the quantization index of the highest hierarchy of the hierarchical quantization method based on the population quantization index appearance frequency supplied from the population quantization index appearance frequency calculation unit 03. In order from the set of, until the specified condition is satisfied, select the set of quantization indexes of the lower hierarchy for the quantization index with a high frequency of appearance (the quantization index of the quantization index of the lower hierarchy is selected). A comparative quantization index set may be selected.
- the specified conditions include, for example, the above-described equal frequency of appearance of the set of quantized indexes, the number of quantized indexes included in the set of quantized indexes (or the depth of the hierarchy), and the invariance of the entire set.
- the degree or a combination thereof may be within a predetermined value range.
- the frequency index appearance frequency equality described above, the number of quantization indexes (or the depth of the hierarchy) included in the quantization index set, the modification invariance of the entire set, and further quantization It may be based on the appearance frequency for each index, the modification invariance for each quantization index, or a combination thereof.
- a minimum value of the appearance frequency equality degree is set, and a quantization index having a high appearance frequency is set until a set of quantization indexes exceeding the minimum value of the appearance frequency uniformity degree appears in order from the quantization index set of the highest hierarchy.
- the quantization index set for comparison may be selected by selecting a quantization index set in the lower hierarchy.
- a minimum value of the number of quantization indexes is set, and the appearance frequency is high until a set of quantization indexes exceeding the minimum value of the number of quantization indexes appears in order from the set of quantization indexes in the highest hierarchy.
- a quantization index set in a lower hierarchy may be selected for the quantization index, and a comparison quantization index set may be selected.
- a minimum value of the appearance frequency uniformity degree and a minimum value of the number of quantization indexes are set, and a set of quantization indexes exceeding the minimum value in order from the set of quantization indexes in the highest hierarchy.
- a set of quantization indexes in lower layers may be selected for a quantization index having a high appearance frequency, and a comparison quantization index set may be selected.
- the quantization index having the highest appearance frequency may be, for example, the quantization index having the highest appearance frequency from the set of quantization indexes that are currently candidates.
- the quantization index with the highest appearance frequency in the set of quantization indexes is the quantization index of the lowest hierarchy.
- a set of quantization indexes in the lower hierarchy may be selected for the quantization index having the second highest appearance frequency.
- it is good also as a some quantization index with a high appearance frequency for example, all the quantization indexes exceeding the threshold with a certain appearance frequency
- one of the sets is selected (see FIG. 6 as an example). For example, assuming that the quantization index having a high appearance frequency is the quantization index 4, the quantization index 13, 14, 15 set or the quantization index 16, 17 set is set as the quantization index of the lower hierarchy. You can choose one). Which one is selected includes, for example, the above-described degree of appearance equality of the set of quantization indexes, the number of quantization indexes included in the set of quantization indexes (or the depth of the hierarchy), and the invariance of the entire set.
- the optimal set may be selected based on the number of quantization indexes (or the depth of the hierarchy) and the modification invariance of the entire set.
- the prescribed condition is “the number of quantization indexes is 10 or more”
- a description will be given of a method of selecting a set of quantization indexes in a lower hierarchy with respect to a quantization index having the highest appearance frequency of a set of candidate quantization indexes.
- the quantization index set ⁇ 1, 2, 3, 4, 5 ⁇ of the highest hierarchy is selected as a candidate. Since the number of selected quantization indexes is 5, the specified condition is not satisfied.
- the quantization index with the highest appearance frequency is 2, then the quantization index 8, 9, 10 which is the quantization index of the lower hierarchy is selected for the quantization index 2, and the quantization index set is set.
- the quantization index with the highest appearance frequency is 3, then the quantization index 11 or 12 that is the quantization index of the lower layer is selected for the quantization index 3, and the quantization index set ⁇ 1 , 8, 9, 10, 11, 12, 4, 5 ⁇ are selected as the next candidates. Since the number of selected quantization indexes is 8, the specified condition is not satisfied.
- the quantization indexes 26 and 27 which are quantization indexes in the lower hierarchy are selected for the quantization index 10, and the quantization index set ⁇ 1 , 8, 9, 26, 27, 11, 12, 4, 5 ⁇ are selected as the next candidates. Since the number of selected quantization indexes is 9, the specified condition is not satisfied.
- the quantization index having the highest appearance frequency is 4, a set of quantization indexes in the lower hierarchy is selected for the next quantization index 4.
- a set of quantization indexes 13, 14, and 15 or a set of quantization indexes 16 and 17 can be selected as the quantization index of the lower layer.
- Which one to select is based on, for example, the above-described degree of frequency index set equalization, the number of quantization indexes (or the depth of the hierarchy), the modification invariance of the entire set, or a combination thereof. You may decide.
- the quantization index set 16, 17 is selected, and the quantization index set ⁇ 1, 8, 9, 26, 27, 11, 12, 16, 17, 5 ⁇ is selected as the next candidate. Since the number of selected quantization indexes is 10, in order to satisfy the prescribed condition, the set of selected quantization indexes ⁇ 1, 8, 9, 26, 27, 11, 12, 16, 17, 5 ⁇ is compared with the comparison quantum.
- the indexed index set is based on, for example, the above-described degree of frequency index set equalization, the number of quantization indexes (or the depth of the hierarchy), the modification invariance of the entire set, or a combination thereof. You may decide.
- the quantization index set 16, 17 is selected, and the quantization index set ⁇ 1, 8, 9, 26, 27, 11, 12, 16, 17, 5 ⁇ is selected as the next candidate. Since the number
- FIG. 10 shows a comparative quantization index set when the method of selecting the quantization index set in the lower hierarchy in order from the quantization index set in the highest hierarchy in the hierarchical quantization method described above.
- the structural example (Structural example 1) of the selection part 021A is shown.
- Configuration Example 1 of the comparative quantization index set selection unit 021A includes candidate quantization index set selection means 021A1 and defined condition determination means 021A2.
- Candidate quantization index set selection means 021A1 selects a set of quantization indexes in the highest hierarchy as a set of candidate quantization indexes when a quantization index (upper hierarchy) is not supplied according to the hierarchical quantization method When a quantization index is supplied, the quantization index is replaced with a set of quantization indexes in the lower layer and selected as a set of candidate quantization indexes. Information on the selected candidate quantization index set is supplied to the prescribed condition determination means 021A2. Note that the quantization index input to the candidate quantization index set selection unit 021A1 is information fed back from the specified condition determination unit 021A2.
- the specified condition determining unit 021A2 determines whether the set of candidate quantized indexes indicated by the information of the candidate quantized index set supplied from the candidate quantized index set selecting unit 021A1 satisfies a predetermined specified condition. To do. If the specified condition is satisfied, the set of candidate quantization indexes is output as a comparison quantization index set. When the prescribed condition is not satisfied, the appearance frequency is high based on the population quantization index appearance frequency supplied from the population quantization index appearance frequency calculation means 03 from the set of candidate quantization indexes. The quantization index is obtained, and the obtained quantization index is supplied (feedback) to the candidate quantization index set selection unit 021A1.
- the conditions specified here are as described above.
- the quantization index set in the lowest hierarchy is sequentially set from the quantization index set of the highest hierarchy in the hierarchical quantization method.
- an “appearance frequency” is selected from the currently set of quantized indexes. The case of selecting a “high” quantization index was described. In the following, a method different from the method of selecting the “highly occurring” quantization index will be described.
- the appearance frequency uniformity degree of a newly generated candidate quantization index set is increased (for example, The quantization index that is the largest) may be selected as a quantization index that is replaced with a set of lower-level quantization indexes.
- quantization index 1 is set to the quantization index of the lower layer.
- the quantization index to be replaced with the lower-layer quantization index set is selected as 3
- the new candidate quantization index set may be ⁇ 1, 2, 11, 12, 4, 5 ⁇ . Further, when selecting, the selection may be made in consideration of the number of quantization indexes included in the newly generated candidate quantization index set.
- a quantization that is replaced with a set of lower-level quantization indexes based on the frequency of appearance of lower-level quantization indexes for each quantization index An index may be selected. For example, for each quantization index, the degree of appearance frequency equality of the set of quantization indexes in the lower layer may be calculated, and the quantization index having the largest set in the lower layer may be selected. Referring to FIG.
- the quantization index of the lower layer of each quantization index within the set ⁇ 6, 7 ⁇ , ⁇ 8, 9, 10 ⁇ , ⁇ 11, 12 ⁇ , ⁇ 13, 14, 15 ⁇ or ⁇ 16, 17 ⁇ , ⁇ 18, 19, 20, 21 ⁇ , for example
- the quantization index to be replaced with the quantization index set in the lower hierarchy is selected as 3, and a new
- the candidate quantization index set may be ⁇ 1, 2, 11, 12, 4, 5 ⁇ . Further, when selecting, the selection may be made in consideration of the number of quantization indexes in the set of quantization indexes in the lower layer.
- a quantization index to be replaced with a set of lower-layer quantization indexes in addition to the appearance frequency, for example, (a) the number of lower-layer quantization indexes for each quantization index, (B) The modification invariance of each quantization index, (c) The modification invariance of a quantization index in a lower layer relative to each quantization index, and the like may be selected.
- the example of (a) will be described with reference to FIG. 6 as an example.
- the set of quantization indexes that are currently candidates is ⁇ 1, 2, 3, 4, 5 ⁇ , for example, quantization index 3 and When the frequency of occurrence of 5 is high, the number of quantization indexes (2 and 4 respectively) of each lower layer is compared.
- the quantization index 5 having a larger number of quantization indexes is compared with the quantization index of the lower layer. You may select as a quantization index replaced with a set of. On the contrary, the quantization index 3 having a small number of quantization indexes may be selected as a quantization index to be replaced with a set of quantization indexes in a lower layer.
- the example of (a) will be described with reference to FIG. 6 as an example.
- the set of quantization indexes that are currently candidates is ⁇ 1, 2, 3, 4, 5 ⁇ , for example, quantization index 3 and
- the frequency of occurrence of 5 is high, for example, for the set ⁇ 11, 12 ⁇ and ⁇ 18, 19, 20, 21 ⁇ of the lower-level quantization index for each quantization index, for example, the modification invariant for the entire set Degree (for example, the average value of the modification invariance of each quantization index), and the larger one of quantization indexes 3 and 5 is selected as a quantization index to be replaced with a set of quantization indexes in lower layers. May be.
- any method may be used.
- the method is not limited to the method described here, and any method may be used as long as it selects a quantization index to be replaced with a set of quantization indexes in a lower layer based on the population quantization index appearance frequency.
- FIG. 11 shows the operation in the second embodiment, particularly when the comparative quantization index set selection unit 021A takes the configuration example 1 (the quantization index of the lower layer is compared with the quantization index having a high appearance frequency).
- the image feature quantity extraction device 01 calculates a hierarchical quantization index code of a population image group from the population image group, and supplies the hierarchical quantization index code to the population quantization index appearance frequency calculation unit 03 (step C01).
- the population quantization index appearance frequency calculation means 03 receives the hierarchical quantization index code of the image group of the population supplied from the image feature quantity extraction device 01 as an input, and performs quantization in the image group of the population.
- the appearance frequency for each index is calculated as the population quantization index appearance frequency, and supplied to the comparison quantization index set selection unit 021A (step C02).
- the candidate quantization index set selection unit 021A1 of the comparison quantization index set selection unit 021A first selects a set of quantization indexes in the highest hierarchy of the hierarchical quantization method as a set of candidate quantization indexes. Then, the information of the selected candidate quantization index set is supplied to the specified condition determining means 021A2 (step C03).
- the specified condition determining unit 021A2 determines whether the set of candidate quantized indexes indicated by the candidate quantized index set information supplied from the candidate quantized index set selecting unit 021A1 satisfies a predetermined specified condition.
- the prescribed condition determining unit 021A2 uses the population quantization index appearance frequency supplied from the population quantization index appearance frequency calculating unit 03 from the set of candidate quantization indexes. Based on this, a quantization index having a high appearance frequency is obtained, and the obtained quantization index is supplied (feedback) to the candidate quantization index set selection unit 021A1 (step C05).
- Candidate quantization index set selection unit 021A1 replaces the quantization index (upper layer) supplied from regulation condition determination unit 021A2 with the set of quantization indexes of the lower layer and selects it as a set of candidate quantization indexes Then, the information of the selected candidate quantization index set is supplied to the defining condition determining unit 021A2 (step C06), and the process proceeds to step C04 again.
- step C04 when the specified condition is satisfied, the specified condition determining unit 021A2 outputs the set of candidate quantized indexes as a comparative quantized index set. Then, the process proceeds to step B02 in the flowchart showing the operation of the image feature quantity comparison device 02 in the first embodiment shown in FIG.
- the comparison quantization index set selection unit 021A causes the appearance of the quantization index based on the population quantization index appearance frequency.
- the discrimination ability which is the degree to which different images can be discriminated. This is because, when the number of quantization indexes is the same, if there is a bias in a specific quantization index (when the frequency of occurrence of a specific quantization index increases), the quantization index with a high frequency of occurrence is different by chance. This is because the probability of appearing in common is increased and the discrimination ability is lowered.
- the ability to discriminate the image feature amount is increased, there is an effect that the accuracy of determining the identity of the image based on the calculated identity scale is increased.
- FIG. 12 is a block diagram illustrating a configuration of an image feature amount comparison apparatus of the image identity scale calculation system according to the third embodiment.
- the comparison quantization index set selection unit 021A in the image feature quantity comparison device 02 of the second embodiment further outputs a quantization index weight value.
- the comparison quantization index set selection unit 021B is replaced, and the quantization index comparison unit 023 is further replaced with a quantization index comparison unit 023A to which a quantization index weight value is supplied as an input. Since the other points are the same as in the second embodiment, only the points different from the second embodiment will be described here.
- the comparison quantization index set selection unit 021B adds the population quantum supplied from the population quantization index appearance frequency calculation unit 03. For each quantization index of the selected comparison quantization index set, the weight value is calculated so as to decrease the weight value as the appearance frequency increases, and the calculated quantization index The weight value is supplied to the quantization index comparison unit 023A as the quantization index weight value.
- the reason why the weight value is decreased as the appearance frequency is high is that the quantization index having a high appearance frequency is likely to have the same quantization index in other images, so that it can be distinguished from other images. This is because the possibility of contributing is low.
- the quantization index weight value may be calculated by any method as long as the weight value is calculated to be smaller as the appearance frequency is higher (that is, the quantization index weight value is monotonically decreased with respect to the appearance frequency). Any method that is calculated based on a function may be used.
- the appearance probability for each quantization index may be calculated from the population quantization index appearance frequency, and the quantization index weight value may be a value obtained by subtracting the appearance probability from 1.
- the quantization index comparison unit 023A compares the comparison quantization indexes of the two images (image 1 and image 2) supplied from the comparison quantization index acquisition unit 022 for each corresponding quantization target region.
- the number of quantization target areas with the same quantization index is obtained for each quantization index, and the quantization index weight value supplied from the comparison quantization index set selection unit 021B is applied to them to calculate the identity measure.
- the identity scale may be calculated, for example, by calculating a product of the number of matching quantization target areas obtained for each quantization index and the quantization index weight value and adding them.
- the comparison quantization index set selection unit 021B sets all the quantizations of the hierarchical quantization method by setting the weight value of the quantization index not included in the selected comparison quantization index to 0 (or lower).
- the weight values of the indexes (all of the quantized indexes 1 to 37 in FIG. 6 as an example) may be supplied to the quantized index comparing unit 023A as the quantized index weight values.
- the quantization index comparison unit 023A since the quantization index weight value not included in the selected comparison quantization index is supplied to the quantization index comparison unit 023A as 0 (or as a low value), the quantization index comparison unit 023A includes: Indirectly, the information of the selected comparison quantization index set is supplied. Therefore, the comparison quantization index set selection unit 021B does not need to supply anything to the comparison quantization index acquisition unit 022 as information on the comparison quantization index set, and the comparison quantization index acquisition unit 022 May acquire the quantization indexes of all layers of the hierarchical quantization method for each quantization target region from the input hierarchical quantization index code, and supply the quantization indexes to the quantization index comparison unit 023A.
- the quantization index comparison unit 023A compares the quantization index for each layer for each corresponding quantization target region, and obtains the number of quantization target regions with the same quantization index for each quantization index.
- the identity measure may be calculated by applying the quantization index weight value supplied from the comparison quantization index set selection unit 021B to them.
- the comparison quantization index set selection unit 021B selects the comparison quantization selected based on the population quantization index appearance frequency. Based on the calculated identity measure by calculating a weight value for each quantization index of the index set and calculating the identity measure by applying the weight value calculated by the quantization index comparison unit 023A. There is an effect that the accuracy of the image identity determination can be made higher than the accuracy of the image identity determination based on the identity scale calculated by the second embodiment.
- FIG. 13 is a block diagram illustrating a configuration of an image feature amount comparison apparatus of the image identity scale calculation system according to the fourth embodiment. Referring to FIG. 13, in the fourth embodiment of the present invention, the comparison quantization index set selection unit 021 of the image feature quantity comparison device 02 having the configuration of the first embodiment shown in FIG. The difference is that the comparison quantization index set selection unit 021C is replaced. Since the other points are the same as in the first embodiment, only the points different from the first embodiment will be described here.
- the comparison quantization index set selection unit 021C receives population quantization index modification invariance as population characteristic information.
- the population quantization index modification invariance represents the modification invariance that is the degree that each quantization index of the hierarchical quantization method does not change (quantization index) by various modification processes in the image group of the population. It is.
- each quantization index of the hierarchical quantization method in the image group of the population for example, the image group of the population (or the image group extracted by appropriately sampling the image group of the population) ) Is subjected to various modification processes to create a duplicate image, and a hierarchical quantization index code is calculated from the original image and the duplicate image using the image feature quantity extraction device 01, and quantization is performed in the corresponding quantization target region
- the ratio of matching indexes may be measured for each quantization index, and the ratio of matching for each quantized index may be defined as the modification invariance for each quantization index.
- the comparison quantization index set selection unit 021C compares the quantization index modification invariance for the entire set of selected quantization indexes based on the input population quantization index modification invariance. Select a quantization index set. Information on the selected comparison quantization index set is supplied to the comparison quantization index acquisition unit 022.
- the comparison quantization index set selection unit 021C is the comparison quantization index set selection unit 021 of the first embodiment, and the input population characteristic information is specifically the population quantization index modification invariance. This is a form of the case, and the conditions relating to the selection of the comparison quantization index set are the same as those described in the first embodiment.
- the comparison quantization index set selection unit 021C selects the comparison quantization index set so that the modification invariance of the quantization index over the entire set of selected quantization indexes is high, and thereby various kinds of images are added.
- the robustness which is the degree that the quantization index is not changed by the modification process, can be increased.
- the comparison quantization index set selection unit 021C based on the population quantization index modification invariance, the comparison quantization index so that the modification invariance of the quantization index applied to the entire set of quantization indexes is increased.
- a method of selecting a set for example, a comparative quantization index set may be selected so that a quantization index having a low population quantization index modification invariance is not included.
- the comparison quantization index set selection unit 021C also performs the comparison quantization based on the population quantization index modification invariance so that the modification invariance of the quantization index applied to the entire set of quantization indexes is increased.
- a quantization index modification invariance over the entire quantization index set is calculated as a quantization index set modification invariance
- the quantization index set may be selected as the comparative quantization index set so that the quantization index set modification invariance becomes high based on the quantization index set modification invariance.
- the comparison quantization index set selection unit 021C calculates the quantization index set modification invariance for each of the combinations of the quantization index sets of all candidates, and the quantization index set modification invariance is calculated.
- the set of quantization indexes that is maximized may be selected as a comparative quantization index set.
- a quantization index set modification invariance which is the modification invariance of the quantization index for the entire set of quantization indexes, is calculated based on the modification invariance of each quantization index included in the set of quantization indexes Any method may be used.
- the quantization index set modification invariance may be calculated as an average value of the modification invariance of each quantization index.
- the quantization index set modification invariance may be calculated as the number of quantization indexes exceeding a certain value, or the ratio thereof.
- the comparison quantization index set selection unit 021C includes the number of quantization indexes included in the set of quantization indexes selected as candidates (or the selected quantization index).
- the comparison quantization index set may be selected in consideration of both the depth of the index set hierarchy). This is because the number of quantization indexes (or the depth of the hierarchy) is determined as information independent of the quantization index set modification invariance and as information independent of the population. The higher the number (or the deeper the hierarchy), the higher the discrimination ability, which is the degree to which different images can be discriminated, but the lower the robustness, the degree to which the quantization index does not change due to various modification processing to the image.
- a comparative quantization index set in consideration of both the quantization index set modification invariance and the number of quantization indexes (or the depth of the hierarchy). For example, from the viewpoint of discrimination ability and robustness, an appropriate range of the number of quantization indexes (for example, the number of quantization indexes is set to about 10) is set, and the quantization index set that satisfies the conditions is set. Therefore, a set that increases the quantization index set modification invariance calculated using the above-described method of calculating the quantization index set modification invariance may be selected as the comparison quantization index set.
- the comparison quantization index set selection unit 021C calculates the appearance frequency equality described in the second embodiment in addition to the above-described quantization index set modification invariance (in this case, the comparison quantization index).
- the set selection unit 021C also needs to be supplied with the population quantization index appearance frequency as an input), and considering both the quantization index set modification invariance and the appearance frequency equality, the comparison quantization index A set may be selected.
- the comparison quantization index set selection unit 021C selects, as the comparison quantization index set, a set of quantization indexes that have a high quantization index set modification invariance and a high appearance frequency equality. May be.
- the comparison quantization index set selection unit 021C does not include, for example, a comparison quantization index so that a degree of appearance frequency equality is high and a quantization index having a low population quantization index modification invariance is not included.
- a set may be selected.
- the comparison quantization index set selection unit 021C sequentially sets the quantization index modification invariance supplied as an input until a prescribed condition is satisfied in order from the quantization index set of the highest hierarchy of the hierarchical quantization method. Based on the above, it is possible to select a set of lower-level quantization indexes and select a comparison quantization index set so that the modification invariance of the quantization index over the entire set of quantization indexes is increased. .
- the specified conditions are, for example, the above-described quantization index set modification invariance, the modification invariance for each quantization index, and the number of quantization indexes included in the set of quantization indexes (or the depth of the hierarchy).
- a combination of these values may be within a predetermined range of values.
- the conditions based on the appearance frequency uniformity degree of the quantization index set described in the second embodiment may be used. For example, the minimum value of the quantization index set modification invariance and the minimum value of the number of quantization indexes are set, and the number of quantization indexes exceeds the number of quantization indexes in order from the set of quantization indexes in the highest hierarchy. A set of quantization indexes in a lower hierarchy may be selected until a set of quantization indexes that does not fall below the set modification invariance appears.
- the same conditions as those described in the ⁇ Configuration example 1 of the comparison quantization index set selection unit 021A> in the second embodiment may be used. (Note that if there are a plurality of sets of quantization indexes calculated by a plurality of different quantization methods in a lower hierarchy of a certain quantization index, one of the sets is selected.)
- a method of selecting a lower-level quantization index set based on the population quantization index modification invariance for example, in a set of quantization indexes that are currently candidates, Then, a set of quantization indexes in the lower layer may be selected (the quantization index is replaced with a set of quantization indexes in the lower layer).
- the modification invariance of quantization index 3 is the largest among them.
- the quantization index 3 is selected as a quantization index to replace the set of lower-layer quantization indexes, and a set of candidate quantization indexes is newly set as ⁇ 1, 2, 11, 12, 4, 5 ⁇ may be selected. Further, when selecting, the selection may be made in consideration of the number of quantization indexes in the set of quantization indexes in the lower layer.
- Another method for selecting a lower-level quantization index set based on the population quantization index modification invariance is, for example, from the set of quantization indexes currently in the candidate, By replacing with a set of quantization indexes, the quantization index set modification invariance, which is the modification invariance of the quantization index over the entire set of newly generated candidate quantization index sets, is increased (for example, maximized).
- the quantization index may be selected as a quantization index that is replaced with a set of lower-level quantization indexes. Referring to FIG. 6 as an example, for example, when the set of quantization indexes that are currently candidates is ⁇ 1, 2, 3, 4, 5 ⁇ , quantization index 1 is set to the quantization index of the lower layer.
- the quantization index to be replaced with the set of quantization indexes in the lower layer is set to 3.
- the new candidate quantization index set may be selected to be ⁇ 1, 2, 11, 12, 4, 5 ⁇ . Further, when selecting, the selection may be made in consideration of the number of quantization indexes included in the newly generated candidate quantization index set.
- a quantization index to be replaced with a set of lower-layer quantization indexes may be selected based on the modification invariance of the lower-layer quantization index with respect to the quantization index. For example, for each quantization index, the quantization invariance (for example, the average value) of the quantization index over the entire set of quantization indexes in the lower layer is calculated, and the quantum having the largest set in the lower layer is calculated. An index may be selected. Referring to FIG.
- the quantization index of the lower layer of each quantization index within the set ⁇ 6, 7 ⁇ , ⁇ 8, 9, 10 ⁇ , ⁇ 11, 12 ⁇ , ⁇ 13, 14, 15 ⁇ or ⁇ 16, 17 ⁇ , ⁇ 18, 19, 20, 21 ⁇ , for example
- the quantization index modification invariance for example, average value
- the quantization index set of the lower hierarchy May be selected as 3, and the new candidate quantization index set may be ⁇ 1, 2, 11, 12, 4, 5 ⁇ . Further, when selecting, the selection may be made in consideration of the number of quantization indexes in the set of quantization indexes in the lower layer.
- the comparison quantization index set selection unit 021C when selecting a lower-level quantization index set based on the population quantization index modification invariance, the population described in the second embodiment in addition to the population quantization index modification invariance Based on the quantization index appearance frequency (in this case, the comparison quantization index set selection unit 021C also needs to be supplied with the population quantization index appearance frequency as an input).
- a quantization index to replace with a set of hierarchical quantization indexes may be selected.
- the configuration of the comparison quantization index selection unit 021C in the case of adopting the method is that the specific operation of each means is different except that the comparison quantization index in the second embodiment shown in FIG. It is almost the same as the set selection unit 021A.
- a candidate quantization index set selection unit that selects a set of candidate quantization indexes
- a specified condition determination unit that determines whether or not a set of quantization indexes selected as a candidate satisfies a specified condition
- the comparison quantization index set selection unit 021C selects the quantization selected based on the population quantization index modification invariance.
- an image feature amount (quantum) expressed using the selected comparison quantization index is selected.
- FIG. 14 is a block diagram illustrating a configuration of an image feature amount comparison apparatus of the image identity scale calculation system according to the fifth embodiment.
- the comparison quantization index set selection unit 021C in the image feature quantity comparison device 02 of the fourth embodiment further outputs a quantization index weight value.
- the comparison quantization index set selection unit 021D is replaced, and the quantization index comparison unit 023 is further replaced with a quantization index comparison unit 023A to which a quantization index weight value is supplied as an input. Since the other points are the same as those in the fourth embodiment, only the points different from the fourth embodiment will be described here.
- the comparison quantization index set selection unit 021D selects based on the population quantization index modification invariance supplied as an input. For each quantization index of the comparison quantization index set, the weight value is calculated so that the weight value is increased as the modification invariance is higher, and the calculated value is supplied to the quantization index comparison unit 023A.
- the quantization index weight value may be calculated by any method as long as the weight value is increased as the modification invariance is higher (that is, monotonous with respect to the modification invariance). Any method that is calculated based on an increasing function may be used. For example, the population quantization index modification invariance itself may be used as the quantization index weight value.
- quantization index comparison unit 023A is the same as the quantization index comparison unit 023A in the third embodiment, a description thereof is omitted here.
- the comparison quantization index set selection unit 021D sets the weight values of quantization indexes not included in the selected comparison quantization index to 0 (or lowers), and performs all quantization of the hierarchical quantization method.
- the weight values of the indexes (all of the quantized indexes 1 to 37 in FIG. 6 as an example) may be supplied to the quantized index comparing unit 023A as the quantized index weight values.
- the quantization index comparison unit 023A since the quantization index weight value not included in the selected comparison quantization index is supplied to the quantization index comparison unit 023A as 0 (or as a low value), the quantization index comparison unit 023A includes: Indirectly, the information of the selected comparison quantization index set is supplied. Therefore, the comparison quantization index set selection unit 021D does not need to supply anything to the comparison quantization index acquisition unit 022 as the comparison quantization index set information, and the comparison quantization index acquisition unit 022 May acquire the quantization indexes of all layers of the hierarchical quantization method for each quantization target region from the input hierarchical quantization index code, and supply the quantization indexes to the quantization index comparison unit 023A.
- the quantization index comparison unit 023A compares the quantization index for each layer for each corresponding quantization target region, and obtains the number of quantization target regions with the same quantization index for each quantization index.
- the identity measure may be calculated by applying the quantization index weight value supplied from the comparison quantization index set selection unit 021D to them.
- the comparison quantization index set selection unit 021D selects the comparison quantum selected based on the population quantization index modification invariance. Based on the calculated identity measure by calculating a weight value for each quantization index of the quantization index set and calculating the identity measure by applying the weight value calculated by the quantization index comparison unit 023A. There is an effect that the accuracy of image identity determination can be made higher than the image identity determination accuracy based on the identity scale calculated by the fourth embodiment.
- FIG. 15 is a block diagram illustrating a configuration of an image identity determination system according to the sixth embodiment.
- the sixth embodiment of the present invention is different in that an identity determination unit 04 is added to the configuration of any of the first to fifth embodiments of the present invention.
- it takes one of the configurations of the first to fifth embodiments for other points only the points different from the first to fifth embodiments will be described here.
- the identity determination unit 04 compares the identity measure of the two images (image 1 and image 2) output from the image feature quantity comparison device 02 with a prescribed threshold supplied as input, and the two images are identical. And the determination result is output as the identity determination result. If the identity measure is larger than the threshold, it is determined that the two images are the same. If the identity measure is smaller than the threshold, it is determined that the two images are not the same.
- the threshold value input here may be a different value depending on the comparison quantization index set selected by the comparison quantization index set selection unit 021. Further, the threshold value may be set by learning using a learning image group in advance.
- the image identity is determined using the identity scale calculated by the image identity scale calculation system according to any one of the first to fifth embodiments.
- the identity of the image is determined based on the identity scale having the effects of the first to fifth embodiments.
- the determination of the identity of the image is performed adaptively optimized for each population, and there is an effect that the determination of the identity of the image can be maintained with high accuracy regardless of the population.
- the seventh embodiment is an embodiment using a hierarchical quantization method as shown in FIG. 16 in the image identity scale calculation system of any of the first to sixth embodiments.
- the quantization method A is the quantization method of the highest layer.
- a feature amount relating to the intensity and direction of the gradient (edge) of the image in the quantization target region is calculated, and when the intensity is less than a predetermined amount, the quantization index indicating “no gradient” is set to 9.
- the direction of the dominant gradient is 8 directions (for example, the horizontal right direction of the image is 0 degree and the clockwise direction is 45 degrees, 90 degrees, 135 degrees, 180 degrees, 225 degrees, 270 degrees and 315 degrees). That is, when the dominant gradient direction is quantized to 0 degrees, it is quantized index 1, when it is quantized 45 degrees, it is quantized index 2, and when it is quantized 90 degrees, it is quantized index. 3 and 270 degrees when quantized at 135 degrees, quantized index 4 when quantized at 180 degrees, quantized index 5 when quantized at 225 degrees, and quantized index 6 when quantized at 225 degrees If it is quantized to 315 degrees, it is classified into a quantization index 7. If it is quantized at 315 degrees, it is classified into a quantization index 8.
- the gradient strength and direction at the center of the quantization target region may be calculated.
- the strength and direction of the gradient (edge) are calculated for each of a plurality of small regions in the quantization target region (for example, each pixel in the quantization target region).
- the direction may be quantized into 8 directions, and voted on a histogram having bins corresponding to each of the 8 directions, and the direction having the maximum number of votes may be the dominant gradient direction (when the number of votes is less than the prescribed number). Is classified into a quantization index 9 indicating “no gradient”).
- the dominant gradient direction is quantized in 8 directions, but it may be quantized in 4 directions or 16 directions, and the direction of quantization is arbitrary. Good.
- the quantization method is used as the quantization method of the lower layer.
- Quantization by B is performed.
- quantization is performed based on a secondary characteristic regarding the gradient.
- the quantization method B may be classified into two based on the amount of gradient strength. Also, for example, whether or not there is a second dominant gradient direction with respect to the dominant gradient direction classified in the highest hierarchy (the intensity of the second dominant gradient direction is the specified amount). It may be classified into two based on whether or not there is more. Alternatively, the second dominant gradient direction may be quantized.
- the classification based on the quantization method B is shown as two classifications, but it is not necessary to be two classifications.
- the quantization method B is similarly applied to all of the quantization indexes 1 to 8 as the quantization method of the lower layer, but for example, only when the dominant gradient direction is a specific angle.
- the quantization method B may be applied.
- the quantization method B is used as the quantization method of the lower layer. It may be applied (other quantization indexes 2, 4, 6, and 8 may be the lowest layer).
- the quantization method B is applied to all of the quantization indexes 1 to 8, and the quantization indexes 10, 11, 14, 15 in which the quantization index of the highest hierarchy is 1, 3, 5, 7 18, 19, 22, and 23 may be further provided with quantization methods in lower layers. This is because it is effective to be able to classify in more detail for a quantization index that is generally assumed to be frequently classified.
- quantization by the quantization method C is performed as the quantization method of the lower layer.
- the quantization method C performs quantization based on a feature different from the gradient (edge) of the image.
- the quantization method C may be quantized based on the luminance of the image in the quantization target region.
- N is a numerical value of 2 or more
- region may be quantized based on the average luminance value of the image of a quantization object area
- the quantization method C may perform quantization based on the color information of the image in the quantization target region.
- a feature related to the hue may be extracted from the image of the quantization target region and quantized based on the angle of the hue.
- the classification according to the quantization method C is shown as three classifications, but it is not necessary to be three classifications.
- the quantization method of the highest layer is based on the direction of the gradient (edge) in general. This is because the characteristic regarding the direction is robust to various modification processes such as filter processing such as image compression and blurring, brightness adjustment, and color tone adjustment. Therefore, for quantization indexes 1 to 8 in which a dominant gradient direction exists, the quantization method of the lower layer is also based on gradient (edge) which is a robust feature (gradient strength, second Dominant gradient direction). Only when it is classified into the quantization index 9 indicating no gradient, gradient information cannot be used as a quantization method in its lower layer (because there is no gradient), and therefore features different from the gradient (luminance, color information, etc.) ) Based on quantization.
- FIG. 17A shows an example of a quantization target region set on the image.
- each region obtained by dividing the image into four parts in the vertical and horizontal directions is set as one quantization target region.
- FIG. 17B shows the quantization indexes extracted from the respective quantization target areas in a predetermined order.
- the hierarchical quantization index code can be stored in the storage means.
- the predetermined order may be arbitrary as long as it is determined in advance.
- the quantization target region (1), the quantization target region (2), the quantization target region (3), and the quantization target region (4). Are arranged in the order.
- the number of quantization target areas per image is variable, the number of quantization target areas per image is stored in the head portion as shown in FIG. 17C, and then each quantization target
- the hierarchical quantization index code can be stored in the storage unit using a data structure in which the quantization indexes extracted from the region are arranged in a predetermined order.
- Each quantization index may include only the quantization index of the lowest layer or may include the quantization index of all layers. For example, taking the hierarchical quantization method shown in FIG. 16 as an example, when the quantization index 11 is obtained for the quantization target region (1), only the quantization index 11 is stored in the quantization index (1). Alternatively, the quantization index 1 and the quantization index 11 may be stored.
- FIG. 18 is a block diagram showing the configuration of the image identity scale calculation system in the present embodiment.
- an outline of the above-described image identity scale calculation system will be described.
- the image identity scale calculation system is a system for calculating an identity scale indicating the degree to which two images are identical, and includes an image feature quantity comparison unit 2A. .
- the image feature quantity comparison unit 2A uniquely identifies a quantization index of a plurality of hierarchies calculated by hierarchical quantization for each quantization target area of two images in accordance with a predetermined hierarchical quantization method.
- a hierarchical quantization index code that is a coding format that can be input, and based on separately provided information, a set of quantization indexes to be used for comparison is selected as a comparison quantization index set, and the comparison quantization index set Is used to compare the hierarchical quantization index codes of the two images and calculate the identity measure of the two images.
- the image identity scale calculation system employs a configuration in which the separately given information is population characteristic information indicating characteristics of a population to which both or at least one of the two images belongs.
- the hierarchical quantization method for each quantization target region of the two images, quantization is performed hierarchically to calculate a quantization index of a plurality of layers, A configuration is adopted in which image feature quantity extraction means for outputting the hierarchical quantization index code, which is an encoding format that can uniquely specify the quantization index of each hierarchy of each quantization target region, is provided.
- the image feature amount extraction unit corresponds to the supplied quantization index when an upper layer quantization index is supplied as feedback according to the hierarchical quantization method for each quantization target region.
- a next layer quantization method selection unit that selects a quantization method of the highest layer, and the selected quantum for each quantization target region.
- a feature amount extracting means for extracting a feature amount used by the quantization method from the input image, and for each quantization target region, the extracted feature amount is quantized according to the selected quantization method.
- a quantization index is calculated, and if the calculated quantization index is not the lowest hierarchy, the quantization index is fed to the next hierarchy quantization method selection means.
- Quantization index calculation means to be supplied as a feedback, and when the quantization index of each hierarchy is calculated for each quantization target area, the quantization index of each hierarchy of each quantization target area
- a hierarchical quantization index code output unit that calculates and outputs a certain hierarchical quantization index code is adopted.
- the image feature amount comparison means includes a comparison quantization index set selection unit that selects a set of quantization indexes used for comparison as a comparison quantization index set based on the population characteristic information, and the two From each of the hierarchical quantization index codes of the image, for each of the quantization target areas, for each image, the hierarchical quantization index code uniquely specified by each hierarchical quantization index for the above comparison
- the comparison quantization index acquisition means for acquiring the quantization index included in the quantization index set as a comparison quantization index, and the comparison quantization index of the two images are compared for each corresponding quantization target region. Based on the number of quantization target regions having the same quantization index, the same scale of the two images That and a quantization index comparing means for calculating a block diagram showing a configuration.
- the image identity scale calculation system adopts a configuration in which the quantization target area is one or a plurality of local areas of the image.
- the hierarchical quantization method is configured by a quantization method in which a quantization index calculated for various modification processes on an image is less likely to change in a higher hierarchy. The structure is taken.
- the quantization method in each layer of the hierarchical quantization method is a plurality of images classified by the quantization method in a general image that does not assume a specific population.
- a configuration is adopted in which the quantization index is composed of quantization methods that are equally classified.
- the image identity scale calculation system adopts a configuration in which the population is an image group included in a classification to which both or at least one of the two images belongs.
- the image identity scale calculation system adopts a configuration in which the population is an image group included in a database to which both or at least one of the two images belongs or a specific subset thereof.
- both or at least one of the two images is a frame of a moving image
- the population includes a moving image or a moving image to which both or at least one of the two images belongs.
- a configuration is adopted in which the frame group is included in the partial section.
- the population characteristic information includes image identity based on an identity scale calculated using a set of selected quantization indexes in the image group of the population.
- the information is correlated with the accuracy of the determination, and the comparison quantization index set is selected so that the accuracy of the determination of the identity of the image based on the calculated identity scale is increased. .
- the population characteristic information has a correlation with an identification capability, which is a degree of identifying a different image included in the set of quantization indexes selected in the image group of the population. It is a certain information and adopts a configuration in which the comparison quantization index set is selected so as to increase the discrimination capability.
- the degree to which the quantization index does not change due to various modification processes to the image included in the set of quantization indexes in the image group of the population is selected so that the robustness is high, and the information is correlated with the robustness.
- the population characteristic information is calculated from a hierarchical quantization index code of the image group of the population output by the image feature quantity extraction unit.
- Population quantization index appearance frequency which is the appearance frequency of each quantization index of the dynamic quantization method
- the comparison quantization index set selection unit determines the quantization index based on the population quantization index appearance frequency. The comparison quantization index set is selected so that the appearance frequency of the.
- the comparison quantization index set selection unit based on the population quantization index appearance frequency, with respect to a quantization index having a high appearance frequency, a quantization index of a lower hierarchy thereof By selecting this set, a configuration is adopted in which a comparative quantization index set is selected.
- the comparison quantization index set selection unit in order from the set of quantization indexes in the highest hierarchy of the hierarchical quantization method, based on the population quantization index appearance frequency. Until a predetermined condition is satisfied, a configuration is adopted in which a set of quantization indexes in a lower hierarchy is selected for a quantization index having a high appearance frequency, and a comparison quantization index set is selected.
- the quantization index set selection unit for comparison does not supply a quantization index as feedback according to the hierarchical quantization method
- the quantization index set of the highest hierarchy is set. Is selected as a candidate quantization index set, and if the quantization index is supplied as feedback, the quantization index is replaced with a set of quantization indexes in the lower layer and selected as a candidate quantization index set Index set selection means and whether or not the candidate quantized index set satisfies a predetermined prescribed condition, and if the prescribed condition is satisfied, the candidate quantized index set is used as a comparative quantized index set Output, If the predetermined condition is not satisfied, a quantization index having a high appearance frequency is fed back to the candidate quantization index set selection unit from the candidate quantization index set based on the population quantization index appearance frequency. It is configured to include a specified condition determining unit to be supplied.
- the comparison quantization index set selection unit generates a comparison quantization index set based on the number of quantization indexes in addition to the population quantization index appearance frequency. The configuration of selecting is adopted.
- the comparison quantization index set selection unit performs, for each quantization index of the selected comparison quantization index set, based on the population quantization index appearance frequency.
- the weight value is calculated as a quantization index weight value so that the weight value is decreased as the appearance frequency is higher, and the quantization index comparison unit converts the comparison quantization index of the two images to the corresponding quantization. Comparing for each target area, the number of quantization target areas with the same quantization index is obtained for each quantization index, and the above-mentioned quantization index weight value is applied to them to calculate the identity measure. Take.
- the population characteristic information indicates that the quantization index of the hierarchical quantization method in the image group of the population does not change the quantization index by various modification processes.
- the quantization index modification invariance shown, and the comparison quantization index set selection unit modifies the quantization index applied to the entire set of quantization indexes selected based on the population quantization index modification invariance
- a configuration is adopted in which a comparative quantization index set is selected so that the invariance is high.
- the comparison quantization index set selection unit sequentially supplies a population quantization index supplied as an input from a set of quantization indexes in the highest hierarchy of the hierarchical quantization method. Based on the modification invariance, a configuration is adopted in which a set of quantization indexes in a lower layer is selected and a comparison quantization index set is selected.
- the comparison quantization index set selection unit generates a comparison quantization index set based on the number of quantization indexes in addition to the population quantization index appearance frequency. The configuration of selecting is adopted.
- the comparison quantization index set selection unit is configured to calculate each quantization index of the selected comparison quantization index set based on the population quantization index modification invariance.
- the weight value is calculated as a quantization index weight value so that the weight value is increased as the modification invariance is higher, and the quantization index comparison unit corresponds to the comparison quantization index of the two images. Comparing for each quantization target region, the number of quantization target regions with the same quantization index is obtained for each quantization index, and the above-mentioned quantization index weight value is applied to them to calculate the identity measure. The structure is taken.
- the image identity scale calculation system further employs a configuration in which it is determined whether or not the two images are the same by comparing the identity scale with a predetermined threshold.
- the image identity scale calculation method which is executed when the above-described image identity scale calculation system operates, calculates an identity scale indicating the degree to which two images are identical.
- a hierarchical quantization index code that is an encoding format that can uniquely identify an index is input, and a set of quantization indexes used for comparison is selected as a comparison quantization index set based on separately provided information, and the comparison
- the hierarchical quantization index codes of the two images are compared using the quantization index set for the two images, and the identity measure of the two images is calculated.
- the image identity scale calculation method adopts a configuration in which the separately given information is population characteristic information indicating characteristics of a population to which both or at least one of the two images belongs.
- the image feature amount extraction unit performs quantization in a hierarchical manner for each quantization target area of the two images according to the hierarchical quantization method.
- the quantization index is calculated, and the hierarchical quantization index code, which is an encoding format that can uniquely identify the quantization index of each layer of each quantization target region, is output to the image feature quantity comparison unit.
- the next-layer quantization method selection unit of the image feature amount extraction unit feeds back the quantization index of the upper layer according to the hierarchical quantization method for each quantization target region. Is selected as the lower layer quantization method corresponding to the supplied quantization index, and if the quantization index is not supplied, the highest layer quantization method is selected, and the image feature is selected.
- the feature amount extraction unit of the amount extraction unit extracts the feature amount used by the selected quantization method for each quantization target region from the input image, and calculates the quantization index of the image feature amount extraction unit.
- the quantization index is supplied as feedback to the next hierarchy quantization method selection means, and the hierarchical quantization index code output means of the image feature quantity extraction means is quantized.
- a hierarchical quantization index code that is a coding format that can uniquely identify the quantization index of each layer of each quantization target region is calculated and output. The configuration is taken.
- the comparison quantization index set selection unit of the image feature amount comparison unit generates a set of quantization indexes used for comparison based on the population characteristic information. An index set is selected, and a comparison quantization index acquisition unit of the image feature amount comparison unit is provided for each quantization target region for each image from each of the hierarchical quantization index codes of the two images. The quantization index included in the comparison quantization index set is acquired as a comparison quantization index from among the quantization indexes of each layer uniquely identified by the hierarchical quantization index code, and the image feature amount comparison is performed. Means for comparing the quantization index for comparison of the two images with a corresponding amount. Reduction compared for each target area, based on the number of quantized regions quantization indexes are matched to calculate the identity scale of the two images, a configuration called.
- the above-described image identity scale calculation system can be realized by incorporating a program into a computer.
- the program according to another aspect of the present invention uses a computer that calculates an identity measure indicating a degree of similarity between two images according to a predetermined hierarchical quantization method.
- the hierarchical quantization index code which is a coding format that can uniquely identify the quantization index of multiple hierarchies calculated by hierarchical quantization for each quantization target area of Then, a set of quantization indexes used for comparison is selected as a comparison quantization index set, the hierarchical quantization index codes of the two images are compared using the comparison quantization index set, and the two images are compared. It is configured to function as an image feature amount comparison unit for calculating the identity measure of
- the program adopts a configuration in which the separately given information is population characteristic information indicating characteristics of a population to which both or at least one of the two images belongs.
- the computer further performs quantization in a hierarchical manner for each quantization target area of the two images according to the hierarchical quantization method to calculate a quantization index of a plurality of layers.
- the image processing apparatus functions as an image feature amount extraction unit that outputs the above-described hierarchical quantization index code, which is an encoding format that can uniquely specify the quantization index of each layer of each quantization target region.
- the image feature quantity extraction unit supplies a quantization index of an upper layer as feedback according to the hierarchical quantization method for each quantization target region, the supplied quantization If the quantization method of the lower layer corresponding to the index is selected and the quantization index is not supplied, the next layer quantization method selection means for selecting the quantization method of the highest layer, and the quantization target region Feature quantity extraction means for extracting the feature quantity used by the selected quantization method from the input image; and for each quantization target area, the extracted feature quantity is quantized according to the selected quantization method. If the calculated quantization index is not the lowest hierarchy, the quantization index is converted to the next hierarchy quantum.
- Quantization index calculation means to be supplied as feedback to the method selection means, and when the quantization index of each hierarchy is calculated for each quantization target area, the quantization index of each hierarchy of each quantization target area can be uniquely identified
- a hierarchical quantization index code output unit that calculates and outputs a hierarchical quantization index code that is an encoding format is adopted.
- the image feature amount comparison unit includes a comparison quantization index set selection unit that selects a set of quantization indexes used for comparison as a comparison quantization index set based on the population characteristic information. From each of the above-mentioned hierarchical quantization index codes of the two images, for each image, for each quantization target region, the hierarchical quantization index code uniquely identifies the quantization index of each layer , A comparison quantization index acquisition means for acquiring a quantization index included in the comparison quantization index set as a comparison quantization index, and a comparison quantization index corresponding to the two images for comparison And based on the number of quantization target regions with the same quantization index, the above One of taking the quantization index comparison means for calculating the identity scale image, a configuration that comprises a.
- the present invention can be used to detect illegal duplication of images or moving images.
- it can be used in a system that detects illegal posting of images / moving images to an image / moving image sharing service on the Internet.
- the link between the original image / moving image and the edited image / moving image can be detected, so it can also be used as a system for managing editing history can do.
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Multimedia (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Library & Information Science (AREA)
- Signal Processing (AREA)
- Data Mining & Analysis (AREA)
- Computer Vision & Pattern Recognition (AREA)
- General Engineering & Computer Science (AREA)
- Databases & Information Systems (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Biology (AREA)
- General Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Software Systems (AREA)
- Computing Systems (AREA)
- Medical Informatics (AREA)
- Health & Medical Sciences (AREA)
- Image Analysis (AREA)
- Compression Or Coding Systems Of Tv Signals (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
Description
そこで、本発明は上記問題点に鑑みて発明されたものであって、その目的は、画像の同一性の判定能力の尺度である識別能力と頑健性とのバランスを調整することのできる画像同一性尺度算出システムを提供することにある。
次に、本発明の第1の実施の形態について図面を参照して詳細に説明する。図3は、第1の実施の形態にかかる画像同一性尺度算出システムの構成を示したブロック図である。図3を参照すると、本発明の第1の実施の形態は、画像特徴量抽出装置01と、画像特徴量比較装置02と、から構成されている。なお、図4は、第1の実施の形態にかかる画像同一性尺度算出システムにおける画像特徴量抽出装置01の具体的な構成を示したブロック図であり、図5は、第1の実施の形態にかかる画像同一性尺度算出システムにおける画像特徴量比較装置02の具体的な構成を示したブロック図である。
次に、図7と図8のフローチャートを利用して、第1の実施の形態の動作を説明する。
次に、本発明の第1の実施の形態の効果について説明する。
次に、本発明の第2の実施の形態について図面を参照して詳細に説明する。第2の実施の形態は、第1の実施の形態における画像特徴量比較装置02において、入力される母集団特性情報と、比較用量子化インデックスセット選択部021がより具体的になったものである。第2の実施の形態における画像特徴量抽出装置は、図4に示した第1の実施の形態の画像特徴量抽出装置01と同じである。図9は、第2の実施の形態にかかる画像同一性尺度算出システムの画像特徴量比較装置の構成を示したブロック図である。図9を参照すると、本発明の第2の実施の形態は、図4と図5に示された第1の実施の形態の構成に、母集団量子化インデックス出現頻度算出手段03が加わり、さらに比較用量子化インデックスセット選択部021が比較用量子化インデックスセット選択部021Aに置き換わる点で異なる。なお、それ以外の点については第1の実施の形態と同じであるため、ここでは第1の実施の形態とは異なる点のみを説明する。
また、比較用量子化インデックスセット選択部021Aは、母集団量子化インデックス出現頻度算出手段03から供給される母集団量子化インデックス出現頻度に基づき、階層的量子化方法の最上位階層の量子化インデックスのセットから順に、規定の条件を満たすまで、出現頻度の高い量子化インデックスに対してその下位階層の量子化インデックスのセットを選択していき(その量子化インデックスをその下位階層の量子化インデックスのセットと置き換える)、比較用量子化インデックスセットを選択してもよい。ここで規定の条件は、例えば、上述した、量子化インデックスのセットの出現頻度均等度合い、量子化インデックスのセットに含まれる量子化インデックスの数(あるいは階層の深さ)や、セット全体の改変不変度や、それらの組み合わせが、あらかじめ規定した値の範囲内であるか否か、などであってもよい。また例えば、上述した、量子化インデックスのセットの出現頻度均等度合い、量子化インデックスのセットに含まれる量子化インデックスの数(あるいは階層の深さ)や、セット全体の改変不変度や、さらに量子化インデックスごとの出現頻度や、量子化インデックスごとの改変不変度や、その組み合わせに基づいたものであってもよい。例えば、出現頻度均等度合いの最小値を設定し、最上位階層の量子化インデックスのセットから順に、出現頻度均等度合いの最小値を超える量子化インデックスのセットが現れるまで、出現頻度の高い量子化インデックスに対してその下位階層の量子化インデックスのセットを選択していき、比較用量子化インデックスセットを選択してもよい。また例えば、量子化インデックスの数の最小値を設定し、最上位階層の量子化インデックスのセットから順に、量子化インデックスの数の最小値を超える量子化インデックスのセットが現れるまで、出現頻度の高い量子化インデックスに対してその下位階層の量子化インデックスのセットを選択していき、比較用量子化インデックスセットを選択してもよい。また例えば、出現頻度均等度合いの最小値と、量子化インデックスの数の最小値の2つを設定し、最上位階層の量子化インデックスのセットから順に、各々の最小値を超える量子化インデックスのセットが現れるまで、出現頻度の高い量子化インデックスに対してその下位階層の量子化インデックスのセットを選択していき、比較用量子化インデックスセットを選択してもよい。また出現頻度の高い量子化インデックスは、例えば、現在候補となっている量子化インデックスのセットの中から最も出現頻度の高い量子化インデックスとしてもよい。なお、例えば現在候補となる量子化インデックスのセットが規定の条件を満たしていない段階で、その量子化インデックスのセットの中の最も出現頻度の高い量子化インデックスが、最下位階層の量子化インデックスである場合は、例えば、2番目に出現頻度の高い量子化インデックスに対してその下位階層の量子化インデックスのセットを選択するようにしてもよい。また、現在候補となっている量子化インデックスのセットの中から出現頻度の高い複数の量子化インデックス(例えば、出現頻度がある閾値を超える全ての量子化インデックス)としてもよい。また、出現頻度の高い量子化インデックスの下位階層に、複数の異なる量子化方法によって算出される複数の量子化インデックスのセットが存在する場合は、そのいずれかのセットを選択する(図6を例に説明すると、例えば出現頻度の高い量子化インデックスが量子化インデックス4であるとして、その下位階層の量子化インデックスとして量子化インデックス13、14、15のセット、あるいは量子化インデックス16、17のセットのいずれかを選択できる)。どれを選択するかは、例えば、上述した、量子化インデックスのセットの出現頻度均等度合い、量子化インデックスのセットに含まれる量子化インデックスの数(あるいは階層の深さ)や、セット全体の改変不変度や、その組み合わせに基づいて決定してもよい。例えば、各々の量子化方法によって算出される量子化インデックスのセットで置き換えて候補量子化インデックスセットを生成した場合に、量子化インデックスのセットの出現頻度均等度合いや、量子化インデックスのセットに含まれる量子化インデックスの数(あるいは階層の深さ)や、セット全体の改変不変度、に基づいて最適となるセットを選択してもよい。
次に、第2の実施の形態において、特に比較用量子化インデックスセット選択部021Aが構成例1を取る場合の動作を、説明する。
まず、画像特徴量抽出装置01は、母集団の画像群から、母集団の画像群の階層的量子化インデックス符号を算出し、母集団量子化インデックス出現頻度算出手段03へ供給する(ステップC01)。次に、母集団量子化インデックス出現頻度算出手段03は、画像特徴量抽出装置01から供給される母集団の画像群の階層的量子化インデックス符号を入力とし、母集団の画像群における、量子化インデックスごとの出現頻度を母集団量子化インデックス出現頻度として算出し、比較用量子化インデックスセット選択部021Aに供給する(ステップC02)。次に、比較用量子化インデックスセット選択部021Aの候補量子化インデックスセット選択手段021A1は、まず階層的量子化方法の最上位階層の量子化インデックスのセットを候補の量子化インデックスのセットとして選択し、選択した候補量子化インデックスセットの情報を規定条件判定手段021A2へ供給する(ステップC03)。次に、規定条件判定手段021A2は、候補量子化インデックスセット選択手段021A1から供給される候補量子化インデックスセットの情報が示す候補の量子化インデックスのセットが、あらかじめ定められた規定の条件を満たすか否かを判定する(ステップC04)。規定の条件を満たさない場合は、規定条件判定手段021A2は、その候補の量子化インデックスのセットの中から、母集団量子化インデックス出現頻度算出手段03から供給される母集団量子化インデックス出現頻度に基づいて、出現頻度の高い量子化インデックスを求め、求めた量子化インデックスを候補量子化インデックスセット選択手段021A1へ供給(フィードバック)する(ステップC05)。候補量子化インデックスセット選択手段021A1は、規定条件判定手段021A2から供給される(上位階層の)量子化インデックスを、その下位階層の量子化インデックスのセットと置き換えて候補の量子化インデックスのセットとして選択し、選択した候補量子化インデックスセットの情報を規定条件判定手段021A2へ供給する(ステップC06)そして再度ステップC04へ進む。ステップC04において、規定の条件を満たす場合は、規定条件判定手段021A2は、その候補の量子化インデックスのセットを、比較用量子化インデックスセットとして出力する。そして、図8に示した第1の実施の形態における画像特徴量比較装置02の動作を示すフローチャートにおけるステップB02へ進む。
次に、本発明の第2の実施の形態の効果について説明する。
次に、本発明の第3の実施の形態について図面を参照して詳細に説明する。図12は、第3の実施の形態にかかる画像同一性尺度算出システムの画像特徴量比較装置の構成を示したブロック図である。図12を参照すると、本発明の第3の実施の形態は、第2の実施の形態の画像特徴量比較装置02における比較用量子化インデックスセット選択部021Aが、さらに量子化インデックス重み値を出力する比較用量子化インデックスセット選択部021Bに置き換わり、また量子化インデックス比較手段023が、さらに量子化インデックス重み値が入力として供給される量子化インデックス比較手段023Aに置き換わる点で異なる。それ以外の点については第2の実施の形態と同じであるため、ここでは第2の実施の形態と異なる点のみを説明する。
次に、本発明の第3の実施の形態の効果について説明する。
次に、本発明の第4の実施の形態について図面を参照して詳細に説明する。第4の実施の形態は、第1の実施の形態における画像特徴量比較装置02において、入力される母集団特性情報と、比較用量子化インデックスセット選択部021がより具体的になったものである。第4の実施の形態における画像特徴量抽出装置は、図4に示した第1の実施の形態の画像特徴量抽出装置01と同じである。図13は、第4の実施の形態にかかる画像同一性尺度算出システムの画像特徴量比較装置の構成を示したブロック図である。図13を参照すると、本発明の第4の実施の形態は、図5に示された第1の実施の形態の構成の画像特徴量比較装置02の比較用量子化インデックスセット選択部021が、比較用量子化インデックスセット選択部021Cに置き換わる点で異なる。なお、それ以外の点については第1の実施の形態と同じであるため、ここでは第1の実施の形態とは異なる点のみを説明する。
次に、本発明の第4の実施の形態の効果について説明する。
次に、本発明の第5の実施の形態について図面を参照して詳細に説明する。図14は、第5の実施の形態にかかる画像同一性尺度算出システムの画像特徴量比較装置の構成を示したブロック図である。図14を参照すると、本発明の第5の実施の形態は、第4の実施の形態の画像特徴量比較装置02における比較用量子化インデックスセット選択部021Cが、さらに量子化インデックス重み値を出力する比較用量子化インデックスセット選択部021Dに置き換わり、また量子化インデックス比較手段023が、さらに量子化インデックス重み値が入力として供給される量子化インデックス比較手段023Aに置き換わる点で異なる。それ以外の点については第4の実施の形態と同じであるため、ここでは第4の実施の形態と異なる点のみを説明する。
次に、本発明の第5の実施の形態の効果について説明する。
次に、本発明の第6の実施の形態について図面を参照して詳細に説明する。第6の実施の形態は、第1から第5の実施の形態のいずれかの画像同一性尺度算出システムを用いて、画像が同一であるか否かを判定する画像同一性判定システムに関する。図15は、第6の実施の形態にかかる画像同一性判定システムの構成を示したブロック図である。図15を参照すると、本発明の第6の実施の形態は、本発明の第1から第5の実施の形態のいずれかの構成に、同一性判定手段04が加わる点で異なる。なお、それ以外の点については第1から第5の実施の形態のいずれかの構成をとるため、ここでは第1から第5の実施の形態とは異なる点のみを説明する。
次に、本発明の第6の実施の形態の効果について説明する。
次に、図16を参照して、本発明の第7の実施の形態について説明する。第7の実施の形態は、上記の第1から第6の実施の形態の何れかの画像同一性尺度算出システムにおいて、図16に示すような階層的量子化方法を用いた実施の形態である。図16において、量子化方法Aは、最上位階層の量子化方法である。量子化方法Aは量子化対象領域における画像の勾配(エッジ)の強度と方向に関する特徴量を算出し、強度が規定の量に満たない場合は「勾配無し」を示す量子化インデックスである9に分類し、それ以外の場合は、支配的な勾配の方向を8方向(例えば画像の水平右方向を0度とし、右廻り方向に順に45度、90度、135度、180度、225度、270度、315度とする)に量子化(分類)する。すなわち支配的な勾配の方向が0度に量子化される場合は量子化インデックス1に、45度に量子化される場合は量子化インデックス2に、90度に量子化される場合は量子化インデックス3に、135度に量子化される場合は量子化インデックス4に、180度に量子化される場合は量子化インデックス5に、225度に量子化される場合は量子化インデックス6に、270度に量子化される場合は量子化インデックス7に、315度に量子化される場合は量子化インデックス8に、分類する。
次に、図17を参照して、本発明の第8の実施の形態について説明する。第8の実施の形態は、上記の第1から第7の実施の形態の何れかの画像同一性尺度算出システムにおいて、画像特徴量抽出装置で作成された階層的量子化インデックス符号を記憶する記憶手段におけるデータ構造として図17に示すようなデータ構造を用いた実施の形態である。図17(a)は画像上に設定される量子化対象領域の例を示しており、この例では、画像を縦横に4分割した各々の領域をそれぞれ1つの量子化対象領域としている。このように、1画像当たりの量子化対象領域の数があらかじめ固定されている場合、図17(b)に示すように、各量子化対象領域から抽出された量子化インデックスを所定の順番で配列したデータ構造を用いて、階層的量子化インデックス符号を記憶手段に記憶することができる。ここで、所定の順番はあらかじめ定められていれば任意でよく、例えば、量子化対象領域(1)、量子化対象領域(2)、量子化対象領域(3)、量子化対象領域(4)の順番に配列される。
次に、本発明の第9の実施の形態を図18を参照して説明する。図18は、本実施の形態における画像同一性尺度算出システムの構成を示すブロック図である。なお、本実施の形態は、上述した画像同一性尺度算出システムの概略を説明する。
111…特徴量抽出手段
112…量子化インデックス算出手段
12…画像特徴量比較装置
121…量子化インデックス比較手段
01…画像特徴量抽出装置
011…次階層量子化方法選択手段
012…特徴量抽出手段
013…量子化インデックス算出手段
014…階層的量子化インデックス符号出力手段
02…画像特徴量比較装置
2A…画像特徴量比較手段
021…比較用量子化インデックスセット選択部
021A…比較用量子化インデックスセット選択部
021A1…候補量子化インデックスセット選択手段
021A2…規定条件判定手段
021B…比較用量子化インデックスセット選択部
021C…比較用量子化インデックスセット選択部
021D…比較用量子化インデックスセット選択部
022…比較用量子化インデックス取得手段
023…量子化インデックス比較手段
023A…量子化インデックス比較手段
03…母集団量子化インデックス出現頻度算出手段
04…同一性判定手段
Claims (35)
- 2つの画像が同一である度合いを示す同一性尺度を算出するシステムであって、
あらかじめ定められた階層的量子化方法に従って、前記2つの画像の量子化対象領域ごとに階層的に量子化を行って算出した複数の階層の量子化インデックスが一意に特定できる符号化形式である階層的量子化インデックス符号を入力とし、別途与えられる情報に基づいて、比較に用いる量子化インデックスのセットを比較用量子化インデックスセットとして選択し、前記比較用量子化インデックスセットを用いて前記2つの画像の階層的量子化インデックス符号を比較し、前記2つの画像の同一性尺度を算出する画像特徴量比較手段、
を備えることを特徴とする画像同一性尺度算出システム。 - 前記別途与えられる情報が、前記2つの画像の両方あるいは少なくとも一方が属する母集団の特性を示す母集団特性情報であることを特徴とする請求項1に記載の画像同一性尺度算出システム。
- 前記階層的量子化方法に従って、前記2つの画像の量子化対象領域ごとに、階層的に量子化を行って複数の階層の量子化インデックスを算出し、各量子化対象領域の各階層の量子化インデックスが一意に特定できる符号化形式である前記階層的量子化インデックス符号を出力する画像特徴量抽出手段を備えることを特徴とする請求項2に記載の画像同一性尺度算出システム。
- 前記画像特徴量抽出手段が、
量子化対象領域ごとに、前記階層的量子化方法に従って、上位階層の量子化インデックスがフィードバックとして供給される場合は、該供給された量子化インデックスに対応する下位階層の量子化方法を選択し、量子化インデックスが供給されない場合は、最上位階層の量子化方法を選択する次階層量子化方法選択手段と、
量子化対象領域ごとに、前記選択された量子化方法が用いる特徴量を、前記入力される画像から抽出する特徴量抽出手段と、
量子化対象領域ごとに、前記抽出された特徴量を、前記選択された量子化方法に従って量子化を行って量子化インデックスを算出し、算出した量子化インデックスが最下位階層でない場合に前記量子化インデックスを前記次階層量子化方法選択手段へフィードバックとして供給する量子化インデックス算出手段と、
量子化対象領域ごとに各階層の量子化インデックスが算出されると、各量子化対象領域の各階層の量子化インデックスが一意に特定できる符号化形式である階層的量子化インデックス符号を算出し、出力する階層的量子化インデックス符号出力手段と、
を備えたことを特徴とする請求項3に記載の画像同一性尺度算出システム。 - 前記画像特徴量比較手段が、
前記母集団特性情報に基づいて、比較に用いる量子化インデックスのセットを比較用量子化インデックスセットとして選択する比較用量子化インデックスセット選択部と、
前記2つの画像の階層的量子化インデックス符号の各々から、各々の画像に対して、量子化対象領域ごとに、階層的量子化インデックス符号が一意に特定する各階層の量子化インデックスの中から、前記比較用量子化インデックスセットに含まれる量子化インデックスを比較用量子化インデックスとして取得する比較用量子化インデックス取得手段と、
前記2つの画像の比較用量子化インデックスを、対応する量子化対象領域ごとに比較し、量子化インデックスが一致する量子化対象領域の数に基づいて、前記2つの画像の同一性尺度を算出する量子化インデックス比較手段と、
を備えたことを特徴とする請求項2乃至4の何れか1項に記載の画像同一性尺度算出システム。 - 前記量子化対象領域が、画像の1つまたは複数の局所領域であることを特徴とする請求項2乃至5の何れか1項に記載の画像同一性尺度算出システム。
- 前記階層的量子化方法が、より上位の階層ほど、画像への各種改変処理に対して算出される量子化インデックスが変化しにくい量子化方法で構成されている、ことを特徴とする請求項2乃至6の何れか1項に記載の画像同一性尺度算出システム。
- 前記階層的量子化方法の各階層での量子化方法は、特定の母集団を仮定しない一般的な画像において、その量子化方法で分類される複数の量子化インデックスに対して、均等に近く分類される量子化方法で構成されている、ことを特徴とする請求項2乃至7の何れか1項に記載の画像同一性尺度算出システム。
- 前記母集団が、前記2つの画像の両方あるいは少なくとも一方が属する分類に含まれる画像群である、ことを特徴とする請求項2乃至8の何れか1項に記載の画像同一性尺度算出システム。
- 前記母集団が、前記2つの画像の両方あるいは少なくとも一方が属するデータベースまたはその特定の部分集合に含まれる画像群である、ことを特徴とする請求項2乃至8の何れか1項に記載の画像同一性尺度算出システム。
- 前記2つの画像の両方あるいは少なくとも一方が動画像のフレームであり、前記母集団が、前記2つの画像の両方あるいは少なくとも一方が属する動画像または動画像の部分区間に含まれるフレーム群である、ことを特徴とする請求項2乃至8の何れか1項に記載の画像同一性尺度算出システム。
- 前記母集団特性情報が、前記母集団の画像群において、選択した量子化インデックスのセットを用いて算出される同一性尺度に基づいた画像の同一性の判定の精度と相関がある情報であり、
前記比較用量子化インデックスセットが、算出される同一性尺度に基づいた画像の同一性の判定の精度が高くなるように選択される、
ことを特徴とする請求項2乃至11の何れか1項に記載の画像同一性尺度算出システム。 - 前記母集団特性情報が、前記母集団の画像群において、選択した量子化インデックスのセットが有する、異なる画像を識別できる度合いである識別能力、と相関がある情報であり、
前記比較用量子化インデックスセットが、識別能力が高くなるように選択される、
ことを特徴とする請求項2乃至11の何れか1項に記載の画像同一性尺度算出システム。 - 前記母集団特性情報が、前記母集団の画像群において、選択した量子化インデックスのセットが有する、画像への各種改変処理によって量子化インデックスが変化しない度合いである頑健性、と相関がある情報であり、
前記比較用量子化インデックスセットが、頑健性が高くなるように選択される、
ことを特徴とする請求項2乃至11の何れか1項に記載の画像同一性尺度算出システム。 - 前記母集団特性情報が、前記画像特徴量抽出手段が出力する前記母集団の画像群の階層的量子化インデックス符号から算出される、前記母集団における階層的量子化方法の各量子化インデックスの出現頻度である母集団量子化インデックス出現頻度であり、
前記比較用量子化インデックスセット選択部が、前記母集団量子化インデックス出現頻度に基づいて、量子化インデックスの出現頻度が均等に近づくように比較用量子化インデックスセットを選択する、
ことを特徴とする請求項2乃至11の何れか1項に記載の画像同一性尺度算出システム。 - 前記比較用量子化インデックスセット選択部が、前記母集団量子化インデックス出現頻度に基づき、出現頻度の高い量子化インデックスに対してその下位階層の量子化インデックスのセットを選択していくことで、比較量子化インデックスセットを選択する、
ことを特徴とする請求項15に記載の画像同一性尺度算出システム。 - 前記比較用量子化インデックスセット選択部が、前記母集団量子化インデックス出現頻度に基づき、階層的量子化方法の最上位階層の量子化インデックスのセットから順に、規定の条件を満たすまで、出現頻度の高い量子化インデックスに対してその下位階層の量子化インデックスのセットを選択していき、比較用量子化インデックセットを選択する、
ことを特徴とする請求項15または16に記載の画像同一性尺度算出システム。 - 前記比較用量子化インデックスセット選択部が、
前記階層的量子化方法に従って、量子化インデックスがフィードバックとして供給されない場合は、最上位階層の量子化インデックスのセットを候補量子化インデックスセットとして選択し、量子化インデックスがフィードバックとして供給される場合は、その量子化インデックスをその下位階層の量子化インデックスのセットと置き換えて候補量子化インデックスセットとして選択する候補量子化インデックスセット選択手段と、
前記候補量子化インデックスセットがあらかじめ定められた規定の条件を満たすか否かを判定し、規定の条件を満たす場合は前記候補量子化インデックスセットを比較用量子化インデックスセットとして出力し、規定の条件を満たさない場合は、前記候補量子化インデックスセットの中から、前記母集団量子化インデックス出現頻度に基づいて、出現頻度の高い量子化インデックスを前記候補量子化インデックスセット選択手段へフィードバックとして供給する規定条件判定手段と、
を備えたことを特徴とする請求項15乃至17の何れか1項に記載の画像同一性尺度算出システム。 - 前記比較用量子化インデックスセット選択部が、前記母集団量子化インデックス出現頻度に加えて、量子化インデックスの数に基づいて、比較用量子化インデックスセットを選択する、
ことを特徴とする請求項15乃至18の何れか1項に記載の画像同一性尺度算出システム。 - 前記比較用量子化インデックスセット選択部が、前記母集団量子化インデックス出現頻度に基づき、選択した比較用量子化インデックスセットの各々の量子化インデックスに対して、出現頻度が高いほど重み値を小さくするように重み値を、量子化インデックス重み値として算出し、
前記量子化インデックス比較手段が、前記2つの画像の比較用量子化インデックスを、対応する量子化対象領域ごとに比較し、量子化インデックスが一致する量子化対象領域の数を、量子化インデックスごとに求め、それらに前記量子化インデックス重み値を作用させて、同一性尺度を算出する、
ことを特徴とする請求項15乃至19の何れか1項に記載の画像同一性尺度算出システム。 - 前記母集団特性情報が、前記母集団の画像群において、階層的量子化方法の各量子化インデックスが、各種改変処理によって量子化インデックスが変化しない度合いを示す母集団量子化インデックス改変不変度であり、
前記比較量子化インデックスセット選択部が、前記母集団量子化インデックス改変不変度に基づいて、選択した量子化インデックスのセット全体にかかる量子化インデックスの改変不変度が高くなるように、比較用量子化インデックスセットを選択する、
ことを特徴とする請求項2乃至20の何れか1項に記載の画像同一性尺度算出システム。 - 前記比較用量子化インデックスセット選択部が、階層的量子化方法の最上位階層の量子化インデックスのセットから順に、入力として供給される母集団量子化インデックス改変不変度に基づき、下位階層の量子化インデックスのセットを選択していき、比較用量子化インデックスセットを選択する、
ことを特徴とする請求項21に記載の画像同一性尺度算出システム。 - 前記比較用量子化インデックスセット選択部が、前記母集団量子化インデックス出現頻度に加えて、量子化インデックスの数に基づいて、比較用量子化インデックスセットを選択する、
ことを特徴とする請求項21または22に記載の画像同一性尺度算出システム。 - 前記比較用量子化インデックスセット選択部が、前記母集団量子化インデックス改変不変度に基づき、選択した比較用量子化インデックスセットの各々の量子化インデックスに対して、改変不変度が高いほど重み値を大きくするように重み値を、量子化インデックス重み値として算出し、
前記量子化インデックス比較手段が、前記2つの画像の比較用量子化インデックスを、対応する量子化対象領域ごとに比較し、量子化インデックスが一致する量子化対象領域の数を、量子化インデックスごとに求め、それらに前記量子化インデックス重み値を作用させて、同一性尺度を算出する、
ことを特徴とする請求項21乃至23の何れか1項に記載の画像同一性尺度算出システム。 - 画像が同一であるか否かを判定するシステムであって、
請求項1乃至24の何れか1項に記載の画像同一性尺度算出システムが出力する同一性尺度を、あらかじめ定められた閾値と比較することにより、前記2つの画像が同一であるか否かを判定する、
ことを特徴とする画像同一性判定システム。 - 2つの画像が同一である度合いを示す同一性尺度を算出する方法であって、
画像特徴量比較手段が、あらかじめ定められた階層的量子化方法に従って前記2つの画像の量子化対象領域ごとに階層的に量子化を行って算出した複数の階層の量子化インデックスが一意に特定できる符号化形式である階層的量子化インデックス符号を入力とし、別途与えられる情報に基づいて、比較に用いる量子化インデックスのセットを比較用量子化インデックスセットとして選択し、前記比較用量子化インデックスセットを用いて前記2つの画像の階層的量子化インデックス符号を比較し、前記2つの画像の同一性尺度を算出する、
ことを特徴とする画像同一性尺度算出方法。 - 前記別途与えられる情報が、前記2つの画像の両方あるいは少なくとも一方が属する母集団の特性を示す母集団特性情報であることを特徴とする請求項26に記載の画像同一性尺度算出方法。
- 画像特徴量抽出手段が、前記階層的量子化方法に従って、前記2つの画像の量子化対象領域ごとに、階層的に量子化を行って複数の階層の量子化インデックスを算出し、各量子化対象領域の各階層の量子化インデックスが一意に特定できる符号化形式である前記階層的量子化インデックス符号を前記画像特徴量比較手段に出力する、
ことを特徴とする請求項27に記載の画像同一性尺度算出方法。 - 前記画像特徴量抽出手段の次階層量子化方法選択手段が、量子化対象領域ごとに、前記階層的量子化方法に従って、上位階層の量子化インデックスがフィードバックとして供給される場合は、該供給された量子化インデックスに対応する下位階層の量子化方法を選択し、量子化インデックスが供給されない場合は、最上位階層の量子化方法を選択し、
前記画像特徴量抽出手段の特徴量抽出手段が、量子化対象領域ごとに、前記選択された量子化方法が用いる特徴量を、前記入力される画像から抽出し、
前記画像特徴量抽出手段の量子化インデックス算出手段が、量子化対象領域ごとに、前記抽出された特徴量を、前記選択された量子化方法に従って量子化を行って量子化インデックスを算出し、算出した量子化インデックスが最下位階層でない場合に前記量子化インデックスを前記次階層量子化方法選択手段へフィードバックとして供給し、
前記画像特徴量抽出手段の階層的量子化インデックス符号出力手段が、量子化対象領域ごとに各階層の量子化インデックスが算出されると、各量子化対象領域の各階層の量子化インデックスが一意に特定できる符号化形式である階層的量子化インデックス符号を算出し、出力する、
ことを特徴とする請求項28に記載の画像同一性尺度算出方法。 - 前記画像特徴量比較手段の比較用量子化インデックスセット選択部が、前記母集団特性情報に基づいて、比較に用いる量子化インデックスのセットを比較用量子化インデックスセットとして選択し、
前記画像特徴量比較手段の比較用量子化インデックス取得手段が、前記2つの画像の階層的量子化インデックス符号の各々から、各々の画像に対して、量子化対象領域ごとに、階層的量子化インデックス符号が一意に特定する各階層の量子化インデックスの中から、前記比較用量子化インデックスセットに含まれる量子化インデックスを比較用量子化インデックスとして取得し、
前記画像特徴量比較手段の量子化インデックス比較手段が、前記2つの画像の比較用量子化インデックスを、対応する量子化対象領域ごとに比較し、量子化インデックスが一致する量子化対象領域の数に基づいて、前記2つの画像の同一性尺度を算出する、
ことを特徴とする請求項27乃至29の何れか1項に記載の画像同一性尺度算出方法。 - 2つの画像が同一である度合いを示す同一性尺度を算出するコンピュータを、
あらかじめ定められた階層的量子化方法に従って、前記2つの画像の量子化対象領域ごとに階層的に量子化を行って算出した複数の階層の量子化インデックスが一意に特定できる符号化形式である階層的量子化インデックス符号を入力とし、別途与えられる情報に基づいて、比較に用いる量子化インデックスのセットを比較用量子化インデックスセットとして選択し、前記比較用量子化インデックスセットを用いて前記2つの画像の階層的量子化インデックス符号を比較し、前記2つの画像の同一性尺度を算出する画像特徴量比較手段、
として機能させるためのプログラム。 - 前記別途与えられる情報が、前記2つの画像の両方あるいは少なくとも一方が属する母集団の特性を示す母集団特性情報であることを特徴とする請求項31に記載のプログラム。
- 前記コンピュータを、さらに、前記階層的量子化方法に従って、前記2つの画像の量子化対象領域ごとに、階層的に量子化を行って複数の階層の量子化インデックスを算出し、各量子化対象領域の各階層の量子化インデックスが一意に特定できる符号化形式である前記階層的量子化インデックス符号を出力する画像特徴量抽出手段、
として機能させる請求項32に記載のプログラム。 - 前記画像特徴量抽出手段が、
量子化対象領域ごとに、前記階層的量子化方法に従って、上位階層の量子化インデックスがフィードバックとして供給される場合は、該供給された量子化インデックスに対応する下位階層の量子化方法を選択し、量子化インデックスが供給されない場合は、最上位階層の量子化方法を選択する次階層量子化方法選択手段と、
量子化対象領域ごとに、前記選択された量子化方法が用いる特徴量を、前記入力される画像から抽出する特徴量抽出手段と、
量子化対象領域ごとに、前記抽出された特徴量を、前記選択された量子化方法に従って量子化を行って量子化インデックスを算出し、算出した量子化インデックスが最下位階層でない場合に前記量子化インデックスを前記次階層量子化方法選択手段へフィードバックとして供給する量子化インデックス算出手段と、
量子化対象領域ごとに各階層の量子化インデックスが算出されると、各量子化対象領域の各階層の量子化インデックスが一意に特定できる符号化形式である階層的量子化インデックス符号を算出し、出力する階層的量子化インデックス符号出力手段と、
を備えたことを特徴とする請求項33に記載のプログラム。 - 前記画像特徴量比較手段が、
前記母集団特性情報に基づいて、比較に用いる量子化インデックスのセットを比較用量子化インデックスセットとして選択する比較用量子化インデックスセット選択部と、
前記2つの画像の階層的量子化インデックス符号の各々から、各々の画像に対して、量子化対象領域ごとに、階層的量子化インデックス符号が一意に特定する各階層の量子化インデックスの中から、前記比較用量子化インデックスセットに含まれる量子化インデックスを比較用量子化インデックスとして取得する比較用量子化インデックス取得手段と、
前記2つの画像の比較用量子化インデックスを、対応する量子化対象領域ごとに比較し、量子化インデックスが一致する量子化対象領域の数に基づいて、前記2つの画像の同一性尺度を算出する量子化インデックス比較手段と、
を備えたことを特徴とする請求項32乃至34の何れか1項に記載のプログラム。
Priority Applications (3)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
US13/061,618 US8897566B2 (en) | 2008-09-01 | 2009-07-14 | Image identity scale calculation stystem |
CN200980138142.1A CN102165490B (zh) | 2008-09-01 | 2009-07-14 | 图像同一性尺度计算系统 |
JP2010526507A JP5482655B2 (ja) | 2008-09-01 | 2009-07-14 | 画像同一性尺度算出システム |
Applications Claiming Priority (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
JP2008-223388 | 2008-09-01 | ||
JP2008223388 | 2008-09-01 |
Publications (1)
Publication Number | Publication Date |
---|---|
WO2010023808A1 true WO2010023808A1 (ja) | 2010-03-04 |
Family
ID=41720993
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/JP2009/003283 WO2010023808A1 (ja) | 2008-09-01 | 2009-07-14 | 画像同一性尺度算出システム |
Country Status (4)
Country | Link |
---|---|
US (1) | US8897566B2 (ja) |
JP (1) | JP5482655B2 (ja) |
CN (1) | CN102165490B (ja) |
WO (1) | WO2010023808A1 (ja) |
Cited By (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102194235A (zh) * | 2010-03-16 | 2011-09-21 | 北京中星微电子有限公司 | 基于梯度方向角的运动检测系统及方法 |
WO2012169965A1 (en) * | 2011-06-08 | 2012-12-13 | Imtt Svenska Ab | Method for comparing images |
JP2021039457A (ja) * | 2019-08-30 | 2021-03-11 | キヤノン株式会社 | 画像処理方法、エッジモデル作成方法、ロボットシステム、および物品の製造方法 |
JP7508206B2 (ja) | 2019-08-30 | 2024-07-01 | キヤノン株式会社 | 画像処理方法、エッジモデル作成方法、ロボットシステム、および物品の製造方法 |
Families Citing this family (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JP5482655B2 (ja) * | 2008-09-01 | 2014-05-07 | 日本電気株式会社 | 画像同一性尺度算出システム |
US9870517B2 (en) * | 2011-12-08 | 2018-01-16 | Excalibur Ip, Llc | Image object retrieval |
ES2907510T3 (es) * | 2012-05-14 | 2022-04-25 | V Nova Int Ltd | Descomposición de datos residuales durante la codificación, decodificación y reconstrucción de señales en una jerarquía escalonada |
US9542619B2 (en) | 2013-03-11 | 2017-01-10 | Yahoo! Inc. | Automatic image piling |
US10031925B2 (en) * | 2014-10-15 | 2018-07-24 | Thinkcx Technologies, Inc. | Method and system of using image recognition and geolocation signal analysis in the construction of a social media user identity graph |
US10678828B2 (en) * | 2016-01-03 | 2020-06-09 | Gracenote, Inc. | Model-based media classification service using sensed media noise characteristics |
JP6566118B2 (ja) * | 2016-03-15 | 2019-08-28 | 日本電気株式会社 | 電子データ検査システム、電子データ検査方法、および電子データ検査用プログラム |
TWI781416B (zh) * | 2019-06-14 | 2022-10-21 | 弗勞恩霍夫爾協會 | 具有基於尺度之改良變換之編碼器、解碼器、方法及電腦程式 |
US12022086B2 (en) * | 2022-08-24 | 2024-06-25 | Synamedia Vividtec Holdings, Inc. | Content-adaptive encoder configuration |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JPS63263983A (ja) * | 1987-04-22 | 1988-10-31 | Matsushita Electric Ind Co Ltd | 階層型符号復号化装置 |
JP2001312514A (ja) * | 2000-03-18 | 2001-11-09 | Hynix Semiconductor Inc | ベクトル記述子表現並びにマルチメディアデータ検索装置及びその方法 |
Family Cites Families (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
KR0132894B1 (ko) | 1992-03-13 | 1998-10-01 | 강진구 | 영상압축부호화 및 복호화 방법과 그 장치 |
US5946044A (en) * | 1995-06-30 | 1999-08-31 | Sony Corporation | Image signal converting method and image signal converting apparatus |
JP3585703B2 (ja) * | 1997-06-27 | 2004-11-04 | シャープ株式会社 | 画像処理装置 |
JP4721111B2 (ja) | 2005-11-24 | 2011-07-13 | 富士ゼロックス株式会社 | 画像処理装置、画像処理システム、画像処理プログラムおよび画像処理方法 |
WO2010023809A1 (ja) * | 2008-09-01 | 2010-03-04 | 日本電気株式会社 | 画像特徴量抽出装置 |
JP5482655B2 (ja) * | 2008-09-01 | 2014-05-07 | 日本電気株式会社 | 画像同一性尺度算出システム |
-
2009
- 2009-07-14 JP JP2010526507A patent/JP5482655B2/ja active Active
- 2009-07-14 CN CN200980138142.1A patent/CN102165490B/zh active Active
- 2009-07-14 WO PCT/JP2009/003283 patent/WO2010023808A1/ja active Application Filing
- 2009-07-14 US US13/061,618 patent/US8897566B2/en active Active
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JPS63263983A (ja) * | 1987-04-22 | 1988-10-31 | Matsushita Electric Ind Co Ltd | 階層型符号復号化装置 |
JP2001312514A (ja) * | 2000-03-18 | 2001-11-09 | Hynix Semiconductor Inc | ベクトル記述子表現並びにマルチメディアデータ検索装置及びその方法 |
Non-Patent Citations (1)
Title |
---|
TAKAYUKI KUROZUMI ET AL.: "Jitsu Kankyo de Shuroku sareta Eizo Danpen o Key to suru Icchi Eizo Tansaku", THE TRANSACTIONS OF THE INSTITUTE OF ELECTRONICS, INFORMATION AND COMMUNICATION ENGINEERS, vol. J90-D, no. 8, 1 August 2007 (2007-08-01), pages 2223 - 2231 * |
Cited By (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102194235A (zh) * | 2010-03-16 | 2011-09-21 | 北京中星微电子有限公司 | 基于梯度方向角的运动检测系统及方法 |
WO2012169965A1 (en) * | 2011-06-08 | 2012-12-13 | Imtt Svenska Ab | Method for comparing images |
JP2021039457A (ja) * | 2019-08-30 | 2021-03-11 | キヤノン株式会社 | 画像処理方法、エッジモデル作成方法、ロボットシステム、および物品の製造方法 |
JP7508206B2 (ja) | 2019-08-30 | 2024-07-01 | キヤノン株式会社 | 画像処理方法、エッジモデル作成方法、ロボットシステム、および物品の製造方法 |
Also Published As
Publication number | Publication date |
---|---|
JP5482655B2 (ja) | 2014-05-07 |
CN102165490A (zh) | 2011-08-24 |
CN102165490B (zh) | 2014-01-22 |
JPWO2010023808A1 (ja) | 2012-01-26 |
US20110170782A1 (en) | 2011-07-14 |
US8897566B2 (en) | 2014-11-25 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
JP5482655B2 (ja) | 画像同一性尺度算出システム | |
JP5445457B2 (ja) | 画像特徴量抽出装置 | |
Wang et al. | Exploring DCT coefficient quantization effects for local tampering detection | |
KR100799557B1 (ko) | 동영상의 비주얼 특징을 이용한 유해 동영상 판별함수 생성및 판별 방법 그리고 그 장치 | |
US8358837B2 (en) | Apparatus and methods for detecting adult videos | |
KR101404401B1 (ko) | 특징량 선택 장치 | |
US7400784B2 (en) | Search of similar features representing objects in a large reference database | |
CN104520875A (zh) | 优选用于搜索和检索目的的从视频内容提取描述符的方法和装置 | |
Pauwels et al. | One class classification for anomaly detection: Support vector data description revisited | |
Ling et al. | Towards perceptually-optimized compression of user generated content (ugc): Prediction of ugc rate-distortion category | |
US20110280487A1 (en) | Image signature extraction device | |
Schettini et al. | Automatic classification of digital photographs based on decision forests | |
Schwamberger et al. | Simple algorithmic modifications for improving blind steganalysis performance | |
CN111611450B (zh) | 跨媒介数据融合方法、装置及存储介质 | |
Youssef et al. | Detecting double and triple compression in HEVC videos using the same bit rate | |
Cheng et al. | High-resolution remote sensing image segmentation based on improved RIU-LBP and SRM | |
Katsenou et al. | Understanding video texture—A basis for video compression | |
KR100811774B1 (ko) | 에지 히스토그램 디스크립터의 특징적 에지 블록을 이용한바이오 영상 검색 방법 및 장치 | |
Schmiedeke et al. | DCT-based features for categorisation of social media in compressed domain | |
Ljubovic et al. | Improving performance of image retrieval based on fuzzy colour histograms by using hybrid colour model and genetic algorithm | |
Subramanian et al. | Anomalous behaviour in loss-gradient based interpretability methods | |
Wang et al. | Steganalysis based on reducing the differences of image statistical characteristics | |
Yu et al. | Diverse Batch Steganography using Model-based Selection and Double-layered Payload Assignment | |
Ndjiki-Nya et al. | Merging MPEG-7 descriptors for image content analysis | |
Das et al. | QPAC: A novel document image compression technique based on ANFIS for calibrated quality preservation |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
WWE | Wipo information: entry into national phase |
Ref document number: 200980138142.1 Country of ref document: CN |
|
121 | Ep: the epo has been informed by wipo that ep was designated in this application |
Ref document number: 09809466 Country of ref document: EP Kind code of ref document: A1 |
|
WWE | Wipo information: entry into national phase |
Ref document number: 2010526507 Country of ref document: JP |
|
WWE | Wipo information: entry into national phase |
Ref document number: 13061618 Country of ref document: US |
|
NENP | Non-entry into the national phase |
Ref country code: DE |
|
122 | Ep: pct application non-entry in european phase |
Ref document number: 09809466 Country of ref document: EP Kind code of ref document: A1 |