WO2010103848A1 - 画像識別子照合装置 - Google Patents
画像識別子照合装置 Download PDFInfo
- Publication number
- WO2010103848A1 WO2010103848A1 PCT/JP2010/001768 JP2010001768W WO2010103848A1 WO 2010103848 A1 WO2010103848 A1 WO 2010103848A1 JP 2010001768 W JP2010001768 W JP 2010001768W WO 2010103848 A1 WO2010103848 A1 WO 2010103848A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- value
- image
- quantized
- quantization
- dimension
- Prior art date
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
-
- 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
- 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/42—Global feature extraction by analysis of the whole pattern, e.g. using frequency domain transformations or autocorrelation
-
- 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/40—Scenes; Scene-specific elements in video content
- G06V20/46—Extracting features or characteristics from the video content, e.g. video fingerprints, representative shots or key frames
Definitions
- the present invention relates to an apparatus for collating images using an image identifier that is a feature amount for identifying an image (determining identity).
- the image identifier is an image feature amount for identifying an image (determining identity). An image identifier extracted from one image is compared with an image identifier extracted from another image, and based on the comparison result, an identity scale (generally, similarity or distance indicating the degree to which the two images are identical) Can be calculated. Further, it is possible to determine whether or not two images are the same by comparing the calculated identity scale with a certain threshold value.
- the two images are the same is not limited to the case where the two images are the same at the level of the image signal (the pixel values of the pixels constituting the image), but the conversion of the compression format (format) of the image, Image size / aspect ratio conversion, image tone adjustment, various image filtering (sharpening, smoothing, etc.), local processing of images (telop overlay, clipping, etc.), image recapturing
- image recapturing The case where one image is a duplicated image of the other image by various modification processes such as. If an image identifier is used, for example, a copy of a moving image that is an image or a collection of images can be detected. Therefore, the image identifier can be applied to an illegal copy detection system for images or moving images.
- FIG. 18 is a diagram illustrating an image identifier extraction method described in Patent Document 1.
- This image identifier is a feature vector of a plurality of dimensions (16 dimensions in FIG. 18).
- An average luminance value is calculated from 32 rectangular areas 244 (of which 16 rectangular areas are drawn in FIG. 18) at predetermined positions in the image 240, and between the paired rectangular areas (see FIG. 18, a pair of rectangular regions are connected by a dotted line 248), and a difference in average luminance value is calculated to obtain a 16-dimensional difference vector 250.
- a composite vector is generated by vector conversion for the difference vector 250, and a 16-dimensional quantization index vector obtained by quantizing each dimension of the composite vector is used as an image identifier.
- An image identifier composed of feature vectors of a plurality of dimensions has a higher level of identification capability that is the degree to which different images can be identified because the smaller the correlation between dimensions, the greater the amount of information that the feature vector has (smaller redundancy). Become.
- the correlation between dimensions is the degree of similarity of occurrence of dimension features, and mathematically, for example, the phase between probability variables when occurrence of feature quantities of each dimension is a random variable. It is a value that can be calculated as the number of relations or mutual information. For this reason, it is desirable that the image identifier composed of feature vectors of a plurality of dimensions is designed so that the correlation between dimensions becomes small.
- the image signal (pixel value of pixels constituting the image) has a correlation between the local regions of the image.
- a specific image pattern repeatedly appears (especially when it appears in a regular cycle) (for example, an image of a building window arranged in a lattice pattern, see FIG. 19A) or In the case of an image composed of a specific texture (see FIG. 19B), the correlation between local regions of the image increases.
- an image identifier composed of a feature vector composed of feature amounts extracted from a plurality of local regions of an image is an image with a large correlation between local regions of the image. Since the shape of the local region from which the feature value is extracted in the dimension is the same (in the example of Patent Document 1, a rectangular region having the same shape), the correlation between the dimensions of the extracted feature value becomes large. Therefore, there is a first problem that the identification ability of the image identifier (feature vector) is lowered.
- the same shape means that the size and the angle (tilt or posture) of the region are the same.
- Patent Document 1 describes an image in which a specific image pattern repeatedly appears (see FIG. 19A) or an image composed of a specific texture (see FIG. 19B). Image identifiers such as those described have poor identification capabilities.
- An object of the present invention is to provide an image identifier collation apparatus that solves the above-described problem, that is, the collation accuracy is low when collation using an image identifier having a low discrimination ability, which is a degree capable of discriminating different images. is there.
- An image identifier verification device extracts a region feature amount from each partial region of a plurality of partial region pairs in an image, and the two partial regions that form a pair for each partial region pair. Quantize the difference value of the region feature value and quantize it to a specific quantization value if the absolute value of the difference value is smaller than a specified value.
- the image identifier of the first image and the image identifier of the second image, which are generated by the generation method, are set to be the specific quantization value. Collating means for collating in a method that reduces the element weight is provided.
- the present invention is configured as described above, it is possible to accurately collate the first image and the second image by using an image identifier having a high identification capability that is a degree capable of identifying different images. In particular, this effect is remarkable for an image having a large correlation between local regions of the image.
- FIG. 1 It is a figure which shows the comparison / quantization method information classified by dimension used in the 5th Embodiment and 6th Embodiment of this invention. It is a figure which shows the comparison / quantization method information classified by dimension used in the 5th Embodiment and 6th Embodiment of this invention. It is a figure which shows the comparison / quantization method information classified by dimension used in the 5th Embodiment and 6th Embodiment of this invention. It is a figure which shows the comparison / quantization method information classified by dimension used in the 5th Embodiment and 6th Embodiment of this invention. It is a figure which shows the extraction method of the image identifier described in patent document 1. FIG.
- 5 is a graph illustrating an example of a monotonic non-increasing function f (D).
- D monotonic non-increasing function
- It is a block diagram of the collation means which collates a quantization index vector.
- It is a flowchart which shows the process example of the collation means which collates a quantization index vector.
- It is a flowchart which shows another example of a process of the collation means which collates a quantization index vector.
- It is a flowchart which shows another example of a process of the collation means which collates a quantization index vector.
- It is a flowchart which shows another example of a process of the collation means which collates a quantization index vector.
- FIG. It is a figure which shows the other example of a process of the collation means which collates a quantization index vector. It is a figure which shows an example of the index provided with respect to 1024 blocks formed by dividing
- FIG. It is a figure which shows the area
- the image identifier extraction apparatus applies a feature vector (more specifically, a quantization index vector) composed of a plurality of dimensions to an input image.
- a feature vector (more specifically, a quantization index vector) composed of a plurality of dimensions to an input image.
- This is a system for outputting as an image identifier, and comprises a dimension determining means 1, an extraction area acquiring means 2, an area feature quantity calculating means 3, and a comparing means 4.
- the dimension determining unit 1 determines the dimension of the feature vector to be extracted next and supplies it to the extraction region acquiring unit 2.
- the dimension determining unit 1 sequentially supplies the dimension of the feature vector to be extracted, and the constituent elements after the extraction region acquiring unit 2 extract the feature amount corresponding to the supplied dimension.
- the dimension determining unit 1 may sequentially supply the first dimension to the Nth dimension to the extraction region acquiring unit 2. If all dimensions of the feature vector are finally supplied, the order of the supplied dimensions may be arbitrary. Multiple dimensions may be supplied in parallel.
- the extraction area acquisition means 2 is supplied with the dimension-specific extraction area information as an input separately from the dimension from the dimension determination means 1.
- the dimension-specific extraction area information is information that indicates a pair of a first extraction area and a second extraction area that are associated with each dimension of the feature vector and that extract the feature quantity of that dimension. is there.
- the first and second extraction regions have the following characteristics as essential conditions.
- An essential condition for the first and second extraction regions is that the relative positions of the extraction region pairs are different among the dimensions, and that the combinations of the shapes of the extraction region pairs are different between the dimensions.
- FIG. 2 shows an example of a pair of extraction areas for each dimension indicated by the dimension-specific extraction information that satisfies the above essential conditions.
- the combination of the shapes of the pairs of extraction regions between dimensions is different.
- the different shapes are congruent shapes with different angles (for example, the first extraction region of the first dimension in FIG. 2 and the first extraction region of the seventh dimension), or similar shapes with different sizes ( For example, the first dimension second extraction area and the ninth dimension second extraction area in FIG. 2 are also included.
- the minimum condition is that at least one dimensional pair with a different combination of shapes of the extracted region pairs exists in all dimensions of the feature vector.
- the more the dimensions of the extraction region pairs (combinations) are different from each other the more desirable. This is because the more the dimensions of the pair of extraction regions (combinations) are different from each other, the smaller the correlation between more dimensions of the feature vector and the higher the discrimination ability.
- the shape (combination) of a pair of extraction regions may be different between all dimensions of the feature vector.
- the first extraction region and the second extraction region in a certain dimension do not need to have the same shape as in the ninth dimension in FIG. 2, but are different in shape from other dimensions in FIG. May be. If the shapes of the first extraction region and the second extraction region in each dimension are different, the correlation between the feature amounts extracted from the first extraction region and the second extraction region is small, and the discrimination ability is high. This is desirable. In addition, since the possibility that the first extraction region and the second extraction region become a frequency blind spot with respect to the same frequency at the same time is reduced, the discrimination capability is increased.
- each extraction area is arbitrary.
- an arbitrary complicated shape such as the second extraction region of the sixth dimension in FIG. 2 may be used.
- it may be a line segment or a curve, for example, as in the seventh dimension or the tenth dimension in FIG.
- the extraction area is composed of a plurality of small areas that are not continuous, such as the first extraction area in the eighth dimension, the first and second extraction areas in the eleventh dimension, and the first extraction area in the twelfth dimension. It may be configured.
- an extraction region having an arbitrarily complicated shape the correlation between dimensions of feature amounts extracted therefrom can be reduced, and the discrimination ability can be increased.
- the first extraction region and the second extraction region may partially overlap. Further, either one of the extraction region pairs may be included in the other. In this way, by allowing duplication of pairs of extraction regions, more patterns of extraction region pairs (relative positions and distances) can be obtained, and therefore, the number of patterns that can reduce the correlation between dimensions can be increased. This increases the possibility of higher identification ability.
- the extraction regions may partially overlap between the dimensions as in each dimension shown in FIG. 2.
- the patterns of extraction area pairs that can be taken are limited.
- FIG. 2 by allowing duplication in the extraction areas between dimensions, more patterns of extraction area pairs can be obtained, so that the number of patterns that can reduce the correlation between dimensions can be increased. , The possibility of higher discrimination ability increases. However, if there are too many extracted regions between dimensions, the correlation between dimensions increases, and the discrimination ability decreases, which is not desirable.
- the extraction area should be such that when the extraction areas of all dimensions are integrated, the area where the feature amount is not extracted in the image becomes small (that is, covers almost the entire screen of the image). .
- the extraction areas of all dimensions are integrated, the area from which no feature value is extracted in the image becomes small (that is, covers almost the entire screen of the image). Since more information contained in can be reflected in the feature amount, the discrimination ability can be increased.
- the extraction regions of all dimensions are integrated, it is desirable that the extraction regions have no bias and are obtained uniformly from the entire image. However, if there is a high probability that local processing such as telop superimposition is performed on a specific area, it is desirable that the extraction area is acquired avoiding that area. In addition, since there are generally no image characteristic portions in the peripheral area such as the edge of the image, it is desirable that the extraction area is acquired avoiding the peripheral area.
- the size and relative position (distance, direction) of the extraction region follow a constant distribution (for example, a uniform distribution).
- a constant distribution for example, a uniform distribution.
- the relative position (distance, direction) follows a uniform distribution, so there is no bias in the distance and direction, and there is no concentration in a specific distance or direction, so more diversity can be taken. Because. Also, the closer the relative position, the greater the correlation between the regions, so in order to cancel it, it is desirable that the closer the relative position, the larger the difference in shape.
- the dimension-specific extraction area information may be information in any format as long as the first extraction area and the second extraction area for each dimension can be uniquely identified.
- the dimension-specific extraction area information indicates the same extraction area for images of any size and aspect ratio.
- the information must be in a format that can be obtained.
- the dimension-specific extraction area information describes the position and shape of the extraction area of an image having a certain size and aspect ratio (for example, an image having a horizontal width of 320 pixels and a vertical width of 240 pixels). Also good.
- the image is first resized to the specified size and aspect ratio, and then the position of the extraction region described in the dimension-specific extraction region information -
- the extraction area may be specified according to the shape.
- the extraction area may be specified by converting the position / shape of the extraction area described in the dimension-specific extraction area information in accordance with an input image having an arbitrary size and aspect ratio. .
- Information indicating each extraction area included in the dimension-specific extraction area information constitutes an extraction area for an image having a certain size and aspect ratio (for example, an image having a horizontal width of 320 pixels and a vertical width of 240 pixels), for example. It may be information describing a set of coordinate values of all pixels.
- the information indicating each extraction area included in the dimension-specific extraction area information may be, for example, information in which the position / shape of the extraction area is described as a parameter for an image having a certain size and aspect ratio. For example, when the shape of the extraction region is a rectangle, information describing the coordinate values of the four corners of the rectangle may be used. For example, when the shape of the extraction region is a circle, the coordinate value and radius value of the center of the circle may be used.
- the pseudo random number seed (seed) is used as the extraction area information for each dimension, starting from the seed within the extraction area acquiring means 2, generating a pseudo random number, and generating an extraction area having a different shape according to the random number.
- the four corners of a quadrangle are determined according to a random number.
- the generated extraction area Since the extraction area is determined based on a random number, the generated extraction area has a different shape for each dimension. If the seeds of the pseudo random numbers are the same, the same random number sequence is generated every time (for any image), so the same extraction region is reproduced for different images.
- the extraction region acquisition unit 2 acquires information indicating the first extraction region and the second extraction region corresponding to the dimension supplied from the dimension determination unit 1 from the extraction region information supplied by dimension supplied as an input, and performs extraction. This is supplied to the area representative value calculation means 3.
- the region feature quantity calculation unit 3 is supplied with an image from which an image identifier is to be extracted as an input. Is done.
- the area feature quantity calculation means 3 includes a first area feature quantity calculation means 31 and a second area feature quantity calculation means 32.
- the area feature quantity calculation means 3 uses the first area feature quantity calculation means 31 to indicate the first extraction area supplied from the extraction area acquisition means 2 for each dimension from the image supplied as input. Based on the above, the feature amount of the first extraction region is calculated as the first region feature amount and supplied to the comparison means 4.
- the area feature quantity calculation unit 3 uses the second area feature quantity calculation unit 32 to calculate the second extraction area supplied from the extraction area acquisition unit 2 for each dimension from the image supplied as input. Based on the indicated information, the feature amount of the second extraction region is calculated as the second region feature amount and supplied to the comparison unit 4.
- the region feature amount calculation unit 3 may select the dimension by size as necessary. Resizes the image to the specified size and aspect ratio of the extraction area information.
- the region feature amount calculation means 3 calculates the region feature amount of each extraction region using the pixel value of the pixel group included in each extraction region.
- the pixel value is a value of a signal possessed by each pixel of the image, and is a scalar amount or a vector amount.
- the pixel value is a luminance value (scalar amount).
- the pixel value is a vector quantity representing a color component.
- the color image is an RGB image
- the pixel value is a three-dimensional vector amount of an R component, a G component, and a B component.
- the color image is a YCbCr image
- the pixel value is a three-dimensional vector amount of a Y component, a Cb component, and a Cr component.
- the calculation method in the extraction region of the dimension is constant (the same calculation method for any input image). As long as it is).
- the region feature amount to be calculated may be a scalar amount or a vector amount.
- the pixel value is a scalar quantity such as a luminance value
- the area feature quantity is calculated as the average value, median value, mode value, maximum value, minimum value, etc. of the pixel values included in the extraction area. (Both are scalar quantities).
- the pixel value at the position of P% from the lower order of the permutation sorted in ascending order is, for example, Y (floor (N ⁇ P / 100)), and this value is calculated as the region feature amount of the extraction region.
- Floor () is a function for truncating after the decimal point.
- the region feature amount calculated by applying this formula (Y (floor (N ⁇ P / 100))) to the luminance value of the pixel included in the extraction region is referred to as a “percentile luminance value feature amount”. I will call it.
- the area feature quantity may be calculated by the method described above after first converting them into a scalar quantity by an arbitrary method.
- the region feature amount may be calculated by the method described above after first converting them into a luminance value that is a scalar amount.
- an average vector of pixel values included in the extraction region may be used as the region feature amount.
- an arbitrary calculation such as edge detection or template matching may be performed on the extraction region, and the calculation result may be used as the region feature amount.
- the region feature amount may be a two-dimensional vector quantity representing the edge direction (gradient direction).
- the edge direction may be a scalar amount that represents the degree of similarity with a certain template.
- a color distribution included in the extraction region, an edge direction distribution, and a histogram representing the edge intensity distribution may be calculated as the region feature amount (all are vector amounts).
- various feature quantities defined in the international standard ISO / IEC 15938-3 that is, Dominant Color, Color Layout, Scalable Color, Color Structure, EdgeHistogram, HomogeneousTexture, Strain. Shape 3D, Parametric Motion, Motion Activity and the like may be used.
- the comparison unit 4 compares, for each dimension, the first region feature amount supplied from the region feature amount calculation unit 3 and the second region feature amount, and the quantization result obtained by quantizing the comparison result Output the index.
- the comparison unit 4 outputs a quantization index for each dimension, so that finally a quantization index vector composed of a plurality of dimensions of the quantization index is output.
- the comparison unit 4 can arbitrarily quantize the first region feature amount and the second region feature amount.
- the number of quantization indexes per dimension is also arbitrary.
- the comparison unit 4 compares the magnitudes, and if the first area feature quantity is larger, the quantization index is +1, otherwise In this case, the quantization index may be set to ⁇ 1, and quantization may be performed into binary quantization indexes of +1 and ⁇ 1.
- the quantization index Qn of dimension n can be calculated by the following equation.
- FIG. 3 shows a more detailed configuration diagram of the comparison unit 4 when the comparison unit 4 performs the comparison / quantization based on the above-described equation 1.
- the comparison means 4 is composed of a magnitude comparison means 41 and a quantization means 42.
- the size comparison means 41 compares the values of the first region feature value and the second region feature value.
- the comparison result is supplied to the quantization means 42. That is, the magnitude comparison means 41 compares the magnitudes of Vn1 and Vn2, and quantifies information indicating whether the comparison result is Vn1> Vn2 or Vn1 ⁇ Vn2 as a magnitude comparison result.
- the generating means 42 To the generating means 42.
- the quantization means 42 performs quantization according to the expression 1 based on the magnitude comparison result supplied from the magnitude comparison means 41, and outputs a quantization index. That is, when information indicating that the comparison result is Vn1> Vn2 is supplied, the quantization unit 42 is supplied with information indicating that the quantization index is +1 and the comparison result is Vn1 ⁇ Vn2. , The quantization index is set to -1, and the quantization index is output.
- comparison / quantization method A the comparison / quantization method based on Equation 1 is referred to as comparison / quantization method A.
- the comparison unit 4 determines the first region feature amount and the first feature amount. 2 is regarded as a quantization index 0 indicating that there is no difference.
- the size is compared, and if the first region feature is larger, the quantum The quantization index may be +1, otherwise the quantization index may be ⁇ 1, and the quantization index may be quantized to +1, 0, ⁇ 1.
- the quantization index Qn of dimension n can be calculated by the following equation. .
- FIG. 4 shows a more detailed configuration diagram of the comparison unit 4 when the comparison unit 4 performs the comparison / quantization based on the above-described Expression 2.
- the comparison unit 4 includes a difference value calculation unit 43 and a quantization unit 44.
- the quantization means 44 is supplied with a threshold value, which is information (quantization boundary information) indicating a quantization boundary, which is defined in advance, as an input.
- the difference value calculation unit 43 calculates a difference value between the value of the first region feature value and the value of the second region feature value.
- the calculated difference value is supplied to the quantization means 44. That is, the difference value calculation unit 43 calculates Vn1 ⁇ Vn2 and supplies the value to the quantization unit 44.
- the quantization means 44 is based on the difference value supplied from the difference value calculation means 43 and a threshold value that is information (quantization boundary information) indicating a predetermined quantization boundary supplied as an input. Quantization is performed according to 2 and a quantization index is output. That is, the quantizing means 42 is based on the value of Vn1 ⁇ Vn2 supplied from the difference value calculating means 41 and the threshold value th supplied as input, in the case of
- the quantization index is +1,
- comparison / quantization method B the comparison / quantization method based on Equation 2 is referred to as comparison / quantization method B.
- the quantization is performed into three values based on the difference value, but it may be quantized into a larger number (of levels) of quantization indexes according to the magnitude of the difference value.
- the comparison means 4 has the configuration shown in FIG. 4, and the quantization means 44 has a plurality of threshold values as information (quantization boundary information) indicating the boundaries of quantization of each level specified in advance. Are supplied as inputs.
- a comparison / quantization method that quantizes a quantization index of a plurality of levels of four or more levels based on the difference value and a plurality of threshold values supplied as input is referred to as a comparison / quantization method C. To.
- the difference between the first region feature value and the second region feature value is small (below the prescribed threshold value)
- a quantization index indicating that there is no difference is introduced.
- the feature quantity (quantization index) of the pair of extracted areas with a small difference in area feature quantity is more stable, that is, more robust to various modification processes and noise. can do. Therefore, the difference in features between local regions is small overall, stable even for flat images with little overall change (for example, images of blue sky), that is, robust to various modification processes and noise, An image identifier (quantization index vector) can be output.
- the comparison unit 4 may first convert the vector quantity into a scalar quantity by an arbitrary method and then perform quantization by the method described above ( This comparison / quantization method will be referred to as comparison / quantization method D). Further, for example, a difference vector that is a difference between the vector of the first extraction region and the vector of the second extraction region may be calculated, and the quantization index may be calculated by vector quantization of the difference vector. In this case, for example, a representative vector (centroid vector or the like) for each predetermined quantization index is supplied, and the quantization index having the largest similarity (the smallest distance) between the representative vector and the difference vector is used.
- comparison / quantization method E this comparison / quantization method will be referred to as comparison / quantization method E.
- this comparison / quantization method when the norm of the difference vector is equal to or smaller than a predetermined threshold, the difference between the first area feature quantity and the second area feature quantity is not found.
- a quantization index 0 indicating that there is no difference
- a quantization index indicating that there is no difference may be introduced.
- the quantization index vectors output in the present invention are collated (the quantization index vector extracted from one image is compared with the quantization index vector extracted from another image, the images are identical).
- the identity measure is calculated as the identity measure.
- the identity measure can be compared with a threshold to determine the identity of the image.
- the identity scale can be calculated as follows. First, the quantization index vectors of two images are compared between corresponding dimensions, and the number of dimensions whose “quantization index is not 0” is calculated (this value is assumed to be A). Next, the number of dimensions with the same quantization index is calculated in dimensions where “both quantization indexes are not 0” (this value is B). Then, the similarity is calculated as B / A.
- A 0 (that is, when the quantization index is 0 for all dimensions), the similarity is set to a specified numerical value (for example, 0.5).
- the identity measure (distance) is calculated as C / A. May be.
- A 0 (that is, when the quantization index is 0 for all dimensions)
- the identity measure (distance) is set to a specified numerical value (for example, 0.5).
- Sequentially for each dimension determine whether the quantization index is ⁇ match / non-match ⁇ , calculate (increment) the value of the number of dimensions ⁇ match / non-match ⁇ the quantization index, and sequentially threshold Compare with.
- the calculation can be aborted when the number of dimensions of the quantization index ⁇ matching / non-matching ⁇ exceeds the threshold value (because it is obvious that the number of dimensions exceeds the threshold value even if it is calculated further).
- the identity scale (similarity) is calculated as B / A
- the identity determination threshold value is larger than 0.5 (half)
- the identity scale (distance) is C /.
- the monotonic non-increasing function f (D) of D is a function that satisfies f (D1) ⁇ f (D2) with respect to D1 ⁇ D2.
- a graph illustrating an example of f (D) with respect to D is shown in FIG. 20 (the horizontal axis is D and the vertical axis is f (D)).
- f (D) does not need to be linear with respect to D as in the functions (i) and (ii), and as long as it is a monotonically non-increasing function of D, the functions (iii) and (iv) A non-linear function such as
- the effect of calculating the identity scale as B / f (D) or C / f (D) using an arbitrary monotonic non-increasing function f (D) of D will be described below.
- the quantization index 0 indicates that there is no difference between the values of the region feature values of the two extraction regions (below the threshold value) with reference to Equation 2.
- a quantization index 0 tends to occur frequently (for example, the whole area).
- a dimension whose quantization index is 0 is It is considered that the effectiveness in determining the identity of images is low.
- the quantization index vectors of the two images being compared, and the dimension where the “quantization index is both 0” is less effective in determining the identity by comparing the quantization index vectors of the two images. It can be considered a dimension.
- B and C exclude dimensions where the effectiveness of both “quantization index is 0” and exclude the dimensions where the effectiveness is “both quantization index is 0”. It is calculated as the number (B) or the number of dimensions (C) in which the quantization indexes do not match.
- f (D) normalizes the values of B and C according to the number of dimensions where “both quantization indexes are 0”, that is, the number of dimensions with low effectiveness (the larger D is, the more A is Since it becomes smaller and the values of B and C become smaller, normalization is performed with a function that does not increase monotonously with respect to D).
- the normalization function an arbitrary monotonic non-increasing function, it is possible to adjust (control) the behavior of the identity scale, and to optimize it according to the image database that performs identity determination and the application There is an effect that can be.
- the method of calculating the identity measure described in the previous paragraph excludes dimensions that have low effectiveness, both of which have a quantization index of 0, and has high effectiveness, dimensions of which both have a quantization index that is not 0.
- An identity scale was calculated for However, the monotonous non-decreasing function g (D) with respect to D which is the number of dimensions having “both quantization index 0” is not excluded without completely excluding the dimension having low effectiveness “both quantization index 0”.
- the identity scale may be calculated as (B / f (D)) + g (D).
- B / f (D) in the first half is an identity measure calculated from a dimension having high effectiveness “both quantization index is not 0”, and g (D) in the latter half is low in effectiveness “both quantization. It is an identity measure calculated from the dimension whose index is “0”, and the overall identity measure is defined as the sum thereof.
- the identity scale can be calculated in the form of increasing the weight of B / f (D).
- the identity scale may be calculated as (B / f (D)) + ⁇ ⁇ D using a small weight value ⁇ .
- the identity scale may be calculated as ⁇ ⁇ B + ⁇ ⁇ D.
- the dimension (number) of “both quantization index is 0” and the dimension (number) of “both quantization index is 0” in paragraphs 0062 to 0065 are respectively set to “any one quantization. It may be read as a dimension (number) whose index is 0 and a dimension (number) whose one quantization index is not 0.
- the collation method (calculation method of identity measure) described in paragraphs 0062 to 0066 is described as a case where the quantization index is calculated based on equation 2, the quantization index is calculated based on equation 2.
- the comparison means 4 is not limited to the case where the difference is made between the first area feature quantity and the second area feature quantity, which are the feature quantities of the two extraction areas (the first extraction area and the second extraction area). If a quantization method that introduces a quantization index indicating that there is no difference (the difference is small and the difference is equal to or less than a predetermined threshold value) is used, the method can be applied.
- the quantization index 0 in the paragraphs 0062 to 0066 is set to “quantization representing that there is no difference between the first area feature quantity and the second area feature quantity (the difference is small, the difference is equal to or less than a predetermined threshold value). It can be interpreted as “index”.
- an image identifier matching device having the matching means as a component is configured. can do.
- the collating unit compares the quantization index vector of the first image output from the comparing unit 4 with the quantization index vector of the second image, calculates an identity measure, and outputs it.
- the image identifier verification device can also include identity determination means in addition to verification means.
- the identity determination means compares the sameness measure supplied from the collation means with a predetermined threshold value to determine whether the first image and the second image are the same, and the determination Output the result.
- the extraction area acquisition means 2 acquires information indicating the first extraction area and the second extraction area of the dimension n from the dimension-specific extraction area information supplied as input, and the area feature amount calculation means 3 (Step A2).
- the area feature quantity calculating means 3 calculates a first area feature quantity and a second area feature quantity of dimension n from the image supplied as input, and supplies them to the comparison means 4 (step A3). ).
- the comparison unit 4 compares the first region feature value of dimension n with the second region feature value, quantizes the comparison result, and outputs a quantization index (step A4).
- step A5 it is determined whether or not the output of the quantization index has been completed for all dimensions (that is, whether n ⁇ N is true or false) (step A5).
- the process ends.
- the output of the quantization index is not completed for all dimensions (that is, when n ⁇ N is true)
- the process proceeds to step A6.
- extraction processing is performed in order from dimension 1 to dimension N, but the order is not limited to this and may be arbitrary.
- extraction processing for a plurality of dimensions may be performed in parallel.
- the first effect is that it is possible to increase the identification ability, which is the degree to which different images can be identified, of image identifiers composed of feature vectors composed of a plurality of dimensions. In particular, this effect is remarkable for an image having a large correlation between local regions of the image.
- the second effect is that the discrimination ability does not deteriorate even for an image in which signals are concentrated at a specific frequency.
- the shape of the region from which features are extracted differs between dimensions (the shape of the region is diverse), so even for images where signals are concentrated at a specific frequency, This is because there is no difference in feature quantity between (many) extraction area pairs (dimensions), and it is difficult for occurrence of a reduction in discrimination ability.
- the second embodiment of the present invention is different in that the comparison means 4 in the first embodiment shown in FIG. 1 is replaced with a comparison means 4A shown in detail in FIG. Other than the comparison unit 4A is the same as that of the first embodiment, and thus the description thereof is omitted here.
- the comparison unit 4 ⁇ / b> A includes a difference value calculation unit 43, a quantization boundary determination unit 45, and a quantization unit 44.
- the difference value calculation unit 43 calculates a difference value between the first region feature amount and the second region feature amount supplied from the region feature amount calculation unit 3 for each dimension, and the quantization boundary determination unit 45 To the quantizing means 44.
- the difference value is, for example, a scalar obtained by subtracting the second region feature amount (or vice versa) from the first region feature amount when the region feature amount is a scalar amount (for example, an average value of luminance values). Amount. Further, when the region feature amount is a vector amount, for example, each vector may be converted into a scalar amount by an arbitrary method, and then the difference value of the scalar amount may be obtained. When the region feature amount is a vector amount, a difference vector between the first region feature amount and the second region feature amount may be used as a difference value (vector amount).
- the quantization boundary determination unit 45 determines the quantization boundary based on the distribution of the difference values of all the dimensions. Then, the information on the determined quantization boundary is supplied to the quantization means 44.
- the distribution of difference values in all dimensions is the frequency (probability) of occurrence with respect to the difference value (or difference vector).
- the determination of the quantization boundary means that, when the difference value is quantized, a parameter to be assigned exclusively to the quantization index is determined without omission.
- the difference value is a scalar quantity, for example, a value range (that is, a threshold value) for each quantization index (quantization level) is determined, and the value range (threshold value) is supplied to the quantization unit 43 as quantization boundary information.
- the difference value is a vector quantity
- a parameter for performing vector quantization for example, a representative vector (such as a centroid vector) of each quantization index is determined, and quantization means is used as information on the quantization boundary. 44.
- the quantization range (threshold value) may be determined so that the ratios of the respective quantization indexes to all dimensions are uniform.
- a constant ⁇ is used.
- the quantization index is +1
- the center point of the difference value distribution (the point at which the integrated values of the left and right distributions are equal) is determined as the quantization threshold ⁇ so that the ratios of the quantization indexes +1 and ⁇ 1 are equal. Also good.
- the difference value is a vector quantity
- the ratios of the respective quantization indexes to all dimensions are made uniform based on the distribution of the difference vectors of all dimensions.
- a vector space region allocated to each quantization index may be determined, or a representative vector (such as a centroid vector) of each quantization index when performing vector quantization may be determined.
- a representative vector such as a centroid vector
- the entropy can be increased by equalizing the ratio of the quantization index to all dimensions (that is, eliminating the bias of the quantization index). can do.
- the quantization boundary determination unit 45 determines the quantization boundary so that the ratio of the quantization index to all dimensions is uniform, and the quantization unit 44 performs quantization based on the comparison / quantization method. This is referred to as comparison / quantization method F.
- the quantization boundary determination unit 45 performs the difference when the difference value is a scalar quantity and the ternary quantization is performed according to the above-described equation 2 (quantization index is +1, 0, ⁇ 1).
- Threshold value th when quantizing to quantization index 0 indicating that there is no is determined based on the distribution of difference values in all dimensions.
- the threshold th may be supplied to the quantization means 44 (in the comparison means 4 of FIG. 4 of the first embodiment, this threshold th is defined in advance).
- the absolute values of the difference values of all dimensions are calculated, the absolute values of the calculated difference values are sorted, and a specified ratio (note that the specified ratio is supplied as an input, for example, from the upper or lower order)
- the comparison / quantization method will be referred to as comparison / quantization method G).
- the threshold th may be determined so that the ratios of the quantization indexes of +1, 0, and -1 approach each other instead of the prescribed ratio (this comparison / quantization method is compared with the comparison / quantization method H). I will call it).
- the comparison / quantization method H corresponds to a specific example of the comparison / quantization method F in the case of following Formula 2.
- Floor () is a function for truncating after the decimal point.
- the method in the present embodiment can be compared with the case in which the comparison unit 4 in the first embodiment has the configuration of FIG.
- a predetermined threshold th is supplied as an input, whereas the above-described method in the second embodiment is performed by the quantization boundary determining unit 45.
- the threshold th is adaptively calculated for the image. As described above, the threshold th is fixed in the first embodiment, and the threshold th is adaptively calculated for the image in the second embodiment.
- the threshold th By calculating the threshold th adaptively for the image, the dimension value of the feature vector is biased toward a specific quantization index (when the specific quantization index is Since the appearance probability is high (especially for an image with few undulations), the discrimination ability can be increased. For example, when the fixed threshold th in the first embodiment is used, an image with few undulations has the majority (or all dimensions) of the feature vector having a quantization index of 0. On the other hand, when the adaptive threshold value th in the second embodiment is used, the threshold value th is automatically adjusted to a small value for an image with few undulations. There is no situation where the index becomes zero.
- the quantization unit 44 performs quantization for each dimension based on the difference value for each dimension supplied from the difference value calculation unit 43 and the quantization boundary information supplied from the quantization boundary determination unit 45. , Output quantization index.
- the quantization means 44 is meaningless if the quantization boundary information outputted from the quantization boundary determination means 45 is ignored, and the quantization boundary determination means 45 determines the quantization boundary. It is necessary to follow the quantization method assumed in the above.
- the extraction region acquisition unit 2 acquires information indicating the first extraction region and the second extraction region of the dimension n from the extraction region information by dimension supplied as input, and the extraction region representative value calculation unit 3 (step B2).
- the extracted area representative value calculation unit 3 calculates the first area feature quantity and the second area feature quantity of the dimension n from the image supplied as input, and supplies them to the difference value calculation unit 43. (Step B3).
- the difference value calculation unit 43 calculates a difference value between the first region feature amount and the second region feature amount of the dimension n, and supplies the difference value to the quantization boundary determination unit 45 and the quantization unit 44. (Step B4).
- step B5 it is determined whether or not the processing up to the calculation of difference values for all dimensions has been completed (that is, whether n ⁇ N is true or false) (step B5).
- the process proceeds to step B7. If the processing for all dimensions has not been completed (that is, if n ⁇ N is true), the process proceeds to step B6.
- extraction processing is performed in order from dimension 1 to dimension N, but the order is not limited to this and may be arbitrary.
- the quantization boundary determination means 45 performs quantization based on the distribution of the difference values of all dimensions.
- the boundary is determined, and information on the determined quantization boundary is supplied to the quantization means 44 (step B7).
- the quantization means 44 performs quantization based on the difference value of the dimension n and the quantization boundary supplied from the quantization boundary determination means 45, and outputs a quantization index (step B9).
- step B10 it is determined whether or not the output of quantization indexes for all dimensions has been completed (that is, whether n ⁇ N is true or false) (step B10).
- the process ends.
- the output of the quantization index for all dimensions is not completed (that is, when n ⁇ N is true)
- the process proceeds to step B11.
- quantization processing is performed in order from dimension 1 to dimension N, but the order is not limited to this and may be arbitrary.
- the quantization boundary is calculated adaptively (dynamically) with respect to the image. Different.
- the dimension value of the feature vector is set to a specific quantum for a specific image (for example, a flat image with few undulations).
- the index is biased toward a specific index (the probability of occurrence of a specific quantization index is high) (entropy is low), and a problem arises in that the discrimination ability of these images is reduced.
- the quantization boundary is adaptively (dynamically) calculated for the image, so that the dimension value of the feature vector for any image becomes Since it can suppress that it is biased to a specific quantization index (the appearance probability of a specific quantization index is high), the identification capability can be increased.
- a region feature quantity calculation method acquisition unit 5 is added to the configuration of the first embodiment shown in FIG. Is different in that it is replaced with a region feature amount calculation unit 3A having first and second region feature amount calculation units 31A and 32A. Since the other configuration is the same as that of the first embodiment, the description thereof is omitted here. In addition, although it demonstrated as a combination with 1st Embodiment here, the combination with 2nd Embodiment may be sufficient.
- the area feature quantity calculation method acquisition means 5 is supplied with the dimensions from the dimension determining means 1 and the dimension-specific area feature quantity calculation method information.
- the area feature value calculation method information for each dimension is information indicating a calculation method of the area feature value in the dimension, which is associated with each dimension of the feature vector defined in advance. Is a necessary condition.
- the difference in the region feature amount calculation method includes a case where different parameters (threshold value, etc.) are applied to the same procedure.
- the region feature amount calculation method is, for example, various methods described in the description of the region feature amount calculation means 3 of the first embodiment, and parameters associated therewith.
- the region feature amount calculation method for each dimension indicated by the region feature amount calculation method information for each dimension is such that at least one pair of different dimension of the region feature amount calculation method exists in all dimensions of the feature vector. It is a condition. It is desirable that there are more dimensions in which the region feature quantity calculation methods are different from each other. This is because as the area feature quantity calculation method has more different dimensions, the correlation between more dimensions of the feature vector becomes smaller and the discrimination ability becomes higher. For example, the region feature amount calculation method may be different between all dimensions of the feature vector.
- the format of the information indicating the region feature value calculation method for each dimension may be any format as long as the method for calculating the region feature value is uniquely specified.
- FIG. 9 shows an example of a region feature amount calculation method for each dimension.
- the region feature quantity calculation method differs between dimensions.
- the feature quantity of the scalar quantity and the vector quantity may be mixed (the first, third, fifth, sixth, eighth, ninth, tenth and twelfth dimensions are the scalar quantity, the second quantity). 4, 7 and 11 dimensions are vector quantities).
- the area feature quantity calculation method acquisition means 5 acquires information indicating the area feature quantity calculation method corresponding to the dimension supplied from the dimension determination means 1 from the dimension-specific area feature quantity calculation method information supplied as input. It is supplied to the feature amount calculation means 3A.
- the area feature quantity calculation means 3A is based on information indicating the first extraction area and the second extraction area supplied from the extraction area acquisition means 2 for each dimension from the image supplied as input. According to the information indicating the region feature amount calculation method supplied from the calculation method acquisition unit 5, the feature amount of the first extraction region and the feature amount of the second extraction region are respectively set to the first region feature amount and the second region feature amount. And is supplied to the comparison means 4.
- the dimension of information indicating the extraction area to be supplied and the dimension of information indicating the area feature quantity calculation method are synchronized.
- the dimension (number) of the feature vector is represented by n, and there are a total of N dimensions from 1 to N.
- the extraction region acquisition unit 2 acquires information indicating the first extraction region and the second extraction region of the dimension n from the dimension-specific extraction region information supplied as input, and the region feature amount calculation unit 3A. (Step C2).
- the region feature amount calculation method acquisition unit 5 acquires information indicating the region feature amount calculation method corresponding to the dimension n from the dimension-specific region feature amount calculation method information supplied as an input, and the region feature amount calculation unit Supply to 3A (step C3).
- the area feature quantity calculating means 3A calculates a first area feature quantity and a second area feature quantity of dimension n from the image supplied as input, and supplies them to the comparison means 4 (step C4). ).
- the comparison unit 4 compares the first area feature quantity of the dimension n with the second area feature quantity, quantizes the comparison result, and outputs a quantization index (step C5).
- it is determined whether or not the output of the quantization index has been completed for all dimensions step C6). If the output of the quantization index is completed for all dimensions, the process ends. If the output of the quantization index has not been completed for all dimensions, the process proceeds to step C7.
- extraction processing is performed in order from dimension 1 to dimension N, but the order is not limited to this and may be arbitrary.
- extraction processing for a plurality of dimensions may be performed in parallel.
- step C2 and step C3 may be reversed.
- the comparison method acquisition means 6 is added to the configuration of the first embodiment shown in FIG. 1, and the comparison means 4 is replaced with the comparison means 4B. It is different in point. Since the other configuration is the same as that of the first embodiment, the description thereof is omitted here. In addition, although it demonstrated as a combination with 1st Embodiment here, the combination with 2nd Embodiment and 3rd Embodiment may be sufficient.
- the comparison method acquisition means 6 is supplied with the dimensions from the dimension determination means 1 and the comparison method information for each dimension.
- the comparison / quantization method information for each dimension is information indicating a method for performing quantization by comparing region feature quantities in the dimensions associated with each dimension of the feature vector defined in advance. It is an essential condition that the comparison and quantization methods are different.
- different comparison / quantization methods include the case where different parameters (threshold, number of quantization indexes, etc.) are applied to the same procedure.
- the comparison / quantization method refers to, for example, various comparison / quantization methods described in the description of the comparison unit 4 of the first embodiment, parameters associated therewith (threshold value, number of quantization indexes, etc.), These are various comparison / quantization methods described in the description of the comparison means 4A of the second embodiment, and accompanying parameters (threshold, number of quantization indexes, etc.).
- the comparison / quantization method for each dimension indicated by the comparison / quantization method information for each dimension is such that at least one pair of different dimensions of the comparison / quantization method exists in all dimensions of the feature vector. It is a condition. The more dimensions the comparison and quantization methods are different from each other, the better. This is because as the comparison / quantization method has more different dimensions, the correlation between more dimensions of the feature vector becomes smaller and the discrimination ability becomes higher. For example, the comparison / quantization method may be different between all dimensions of the feature vector.
- the format of the information indicating the comparison / quantization method for each dimension may be any format as long as the method of comparing and quantizing the region feature quantities is uniquely specified.
- FIG. 12 shows an example of a comparison / quantization method for each dimension.
- the comparison / quantization method differs between dimensions.
- different parameters may be set by the same comparison / quantization method as in the third, fifth, and twelfth dimensions.
- the example of the comparison / quantization method for each dimension shown in FIG. 12 corresponds to the example of the region feature value calculation method for each dimension shown in FIG.
- the comparison / quantization method of scalar quantities is shown
- the comparison / quantization method of vector quantities is shown as an example for region feature quantities of vector quantities.
- the comparison method acquisition unit 6 acquires information indicating the comparison / quantization method corresponding to the dimension supplied from the dimension determination unit 1 from the dimension-specific comparison / quantization method information supplied as an input, and sends the information to the comparison unit 4B. Supply.
- the comparison unit 4B compares, for each dimension, the first region feature amount and the second region feature amount supplied from the region feature amount calculation unit 3 and a comparison / quantization method supplied from the comparison method acquisition unit 6. Are compared and quantized according to the information indicating the output, and a quantization index is output.
- the comparison unit 4B may be configured to include both the comparison unit 4 of the first embodiment and the comparison unit 4B of the second embodiment as necessary. .
- the comparison means 4B needs to synchronize the dimension of the supplied region feature quantity with the dimension of the information indicating the comparison / quantization method.
- the extraction area acquisition means 2 acquires information indicating the first extraction area and the second extraction area of the dimension n from the dimension-specific extraction area information supplied as input, and the area feature amount calculation means 3 (Step D2).
- the comparison method acquisition unit 6 acquires information indicating the comparison / quantization method corresponding to the dimension n from the dimension-specific comparison / quantization method information supplied as an input, and supplies the information to the comparison unit 4B (step). D3).
- the area feature quantity calculating means 3 calculates the first area feature quantity of the dimension n and the second area feature quantity from the image supplied as input, and supplies them to the comparison means 4B (step D4). ).
- the comparison unit 4B compares the first area feature quantity of the dimension n and the second area feature quantity, quantizes the comparison result, and outputs a quantization index (step D5).
- it is determined whether or not the output of the quantization index has been completed for all dimensions step D6). If the output of the quantization index is completed for all dimensions, the process ends. If the output of the quantization index has not been completed for all dimensions, the process proceeds to step D7.
- extraction processing is performed in order from dimension 1 to dimension N, but the order is not limited to this and may be arbitrary. In addition to this processing procedure, extraction processing for a plurality of dimensions may be performed in parallel. Furthermore, the order of step D2 and step D3 may be reversed, and step D3 may be executed immediately before step D5.
- the image identifier extraction device of the present invention can be realized by a computer and a program as well as by realizing the functions of the image identifier extraction device in hardware.
- the program is provided by being recorded on a computer-readable recording medium such as a magnetic disk or a semiconductor memory, and is read by the computer at the time of starting up the computer, etc.
- dimensional determination means extraction area acquisition means, area feature quantity calculation means, comparison means, area feature quantity calculation method acquisition means, and comparison method acquisition means.
- the feature vector to be extracted has 300 dimensions (from the first dimension to the 300th dimension).
- each dimension extraction area is composed of quadrangular shapes.
- FIG. 14 shows dimension-specific extraction region information supplied as an input to the extraction region acquisition means 2 in the fifth embodiment.
- FIG. 14 shows the XY coordinate values of the four corners of the extraction area (first extraction area and second extraction area) for each dimension with respect to an image size of 320 pixels wide ⁇ 240 pixels long, which is a prescribed image size. Indicates.
- the extraction area of the first dimension is configured by a rectangle having four corners with coordinate values (262.000,163.000), coordinate values (178.068,230.967), coordinate values (184.594,67.411), and coordinate values (100.662,135.378).
- a first extraction area and a first rectangle composed of four corners with coordinate values (161.000,133.000), coordinate values (156.027,132.477), coordinate values (164.240,102.170), and coordinate values (159.268,101.647) It consists of an extraction area.
- An extraction area for each dimension is an area surrounded by the coordinate values of the four corners of an image normalized to an image size of 320 pixels wide by 240 pixels wide. It is a set of pixels with integer coordinate values included in. However, negative coordinate values included in the region surrounded by the coordinate values at the four corners are not included in the extraction region.
- FIG. 15 shows the dimension-specific region feature value calculation method information supplied as an input to the region feature value calculation method acquisition unit 5 in the fifth embodiment.
- the average value of the luminance values of the pixel groups included in each extraction region is This is a region feature amount.
- FIG. 17 shows dimension-specific comparison / quantization method information supplied as an input to the comparison method acquisition unit 6 in the fifth embodiment.
- the comparison / quantization method B or the comparison / quantization method G is used for each dimension, and the parameter values are different for each dimension.
- the feature vector to be extracted has 300 dimensions (from the first dimension to the 300th dimension).
- the information shown in FIG. 14 is used as dimension-specific extraction region information supplied as an input to the extraction region acquisition means 2 as in the fifth embodiment.
- the information shown in FIG. 17 is used as the dimension-specific comparison / quantization method information supplied as an input to the comparison method acquisition unit 6 as in the fifth embodiment.
- FIG. 16 shows the dimension-specific region feature value calculation method information supplied as an input to the region feature value calculation method acquisition unit 5 in the sixth embodiment.
- the average value of the luminance values of the pixel groups included in the extraction region (the first extraction region and the second extraction region) or the percentile luminance value feature amount is used for each dimension. Even when the same percentile luminance value feature value is used, the feature value is different for each dimension.
- the first dimension is an average value of luminance values of pixels included in the extraction region.
- the fourth dimension is a percentile luminance value feature amount, Y (floor (N ⁇ 20.0 / 100).
- the eighth dimension is a percentile luminance value feature amount, Y (floor (N ⁇ 80.0 / 100). 100).
- the feature vector to be extracted has 325 dimensions (from the first dimension to the 325th dimension).
- each area is configured by a combination of 1024 blocks that are obtained by dividing an image in the vertical direction 32 and the horizontal direction 32.
- an index starting from 0 is assigned to each block in order from the upper left, and an area is described using this index.
- the rectangular area is expressed as ab using the index a of the upper left block and the index b of the lower right block.
- a rectangle composed of four blocks with indexes 0, 1, 32, and 33 is described as 0-33.
- 2-67 is an area formed by connecting a rectangle defined by 0-33 and a rectangle defined by 2-67, that is, block numbers 0, 1, 2, 3, 32, An area composed of 33, 34, 35, 66, and 67 is shown.
- FIG. 28 shows a region corresponding to each dimension of the seventh embodiment by this notation.
- the region type is a grouping (typing) of regions having similar region patterns determined by a combination of relative positions and shapes between the first and second extraction regions.
- first and second extraction regions are both rectangles consisting of 4 vertical blocks and 2 horizontal blocks, or rectangles consisting of 2 vertical blocks and 4 horizontal blocks.
- the relative positional relationship between the first and second extraction regions is viewed, they exist at adjacent positions so that the long sides of the rectangle overlap each other.
- the shape of the first and second extraction regions is a shape in which two squares each having two vertical and horizontal blocks are arranged on a diagonal of 45 degrees or 135 degrees so as to share one vertex.
- the two squares constituting the second area are present at positions immediately adjacent to the left and below the upper left square of the first area. .
- the shapes of the first and second extraction regions are both squares composed of 10 blocks vertically and horizontally. Further, when looking at the relative positional relationship between the first and second extraction regions, they exist at positions separated by an integral multiple of 10 blocks both vertically and horizontally.
- the first and second extraction regions are both squares of 6 blocks in length and width. Further, when looking at the relative positional relationship between the first and second extraction regions, they exist at positions separated by an integral multiple of 6 blocks both vertically and horizontally.
- first and second extraction areas two areas formed by dividing a square area into a central square and two outside thereof are defined as first and second extraction areas.
- the shape of the region is such that the second extraction region is a central square, and the first square is a shape obtained by hollowing out the second extraction region from the entire square. Further, when the relative positional relationship of the regions is viewed, the second extraction region exists at the position of the center hole of the first extraction region.
- the shape of the area is as follows: the first extraction area is 6 blocks long, the rectangle is 10 blocks wide, the second extraction area is 10 blocks long, It is a 6-block rectangle. Further, when the relative positional relationship between the first and second extraction regions is viewed, the center positions are arranged to coincide with each other.
- a rectangle consisting of 4 vertical blocks and 12 horizontal blocks, or a rectangle consisting of 12 vertical blocks and 4 horizontal blocks, is divided into three equal parts. This corresponds to the case where the center square formed and the other two regions are the first and second extraction regions.
- the shape of the region is such that the first extraction region is a shape in which two squares composed of 4 blocks in length and width are arranged 4 blocks apart vertically or horizontally, and the second extraction region is a square composed of 4 blocks in length and width. is there.
- a second extraction area exists between the first extraction areas.
- region types in FIGS. 28-a, 28-b, 28-c, 28-d, 28-e, 28-f, and 28-g are respectively referred to as region type a, region type b, These will be referred to as region type c, region type d, region type e, region type f, and region type g.
- the average luminance value is calculated as the region feature amount, and the feature amount in each dimension is calculated.
- a value extracted by the above-described various extraction methods such as median or maximum value instead of the average of luminance values may be obtained as the region feature amount.
- the index corresponding to the threshold is 37.
- indexes corresponding to threshold values can be obtained for other region types. This is shown in FIG. In this way, when the threshold value is determined for each region type, the probability of occurrence of 0, 1, and ⁇ 1 in each dimension can be made uniform as compared with the case where the threshold value is determined as a whole, and the discrimination ability is improved.
- the quantization may be performed by the other various quantization methods described above.
- a representative value for example, an average value of luminance values of pixels in a block
- an area feature amount is extracted therefrom. You may do it.
- region type has symmetry as a whole. For this reason, even when the right and left of the image are inverted or the image is inverted up and down, by appropriately changing the correspondence and sign of the dimensions, the feature amount extracted from the image that is horizontally or vertically inverted is also used. The image feature amount can be restored. For this reason, it becomes possible to collate with an image that is reversed left and right or up and down.
- collating means for collating the quantization index vector output in the present invention will be described with reference to a block diagram.
- FIG. 21 there is shown a block diagram of the collating means 100 for collating the quantized index vectors output in the present invention.
- the dimension determining means 101 is connected to the quantized value acquiring means 102 and 103 and outputs the determined dimension information.
- the quantized value acquiring unit 102 acquires the quantized index value of the dimension input from the dimension determining unit 101 from the first quantized index vector, and outputs it to the scale calculating unit 104 as the first quantized index value.
- the quantization value acquisition unit 103 acquires the dimension quantization index value input from the dimension determination unit 101 from the second quantization index vector, and outputs it to the scale calculation unit 104 as the second quantization index value.
- the scale calculation means 104 calculates and outputs a scale representing identity from the first and second quantization index values output from the quantized value acquisition means 102 and 103, respectively.
- the collation unit 100 receives a first quantization index vector that is a quantization index vector extracted from the first image and a second quantization that is a quantization index vector extracted from the second image. An index vector is input. The input first and second quantization index vectors are input to the quantized value acquisition means 102 and 103, respectively.
- Dimension information output from the dimension determining means 101 is also input to the quantized value acquisition means 102 and 103.
- the dimension determining unit 101 sequentially outputs information specifying each dimension of the quantization index vector which is an N-dimensional vector.
- the order of output does not necessarily increase one by one from 1 to N, and any order may be used as long as the dimensions from 1 to N are specified without excess or deficiency.
- the quantized value acquisition means 102 and 103 acquire the quantization index value of the dimension specified by the dimension information output from the dimension determination means 101 from the input quantization index vector. Then, the acquired quantization index value is output to the scale calculation means 104.
- the scale calculation unit 104 compares the first quantization index value output from the quantization value acquisition unit 102 with the second quantization index value. This comparison is performed for each dimension, and a similarity measure (or distance measure) between the first and second quantization index vectors is calculated as an identity measure.
- the obtained identity scale value is compared with a predetermined threshold value to determine identity.
- identity scale is a scale representing similarity, it is determined that they are the same when the scale value is equal to or greater than a threshold value.
- identity scale is a scale representing a distance, it is determined that they are the same when the scale value is equal to or smaller than a threshold value.
- FIG. 22 is a flowchart showing the operation of the verification unit 100.
- the dimension (number) of the quantization index vector is represented by n, and there are a total of N dimensions from 1 to N.
- a variable for calculating the similarity is represented by B.
- the quantized value acquisition means 102 and 103 the first quantized index value and the second quantized index value of the dimension n are obtained from the first quantized index vector and the second quantized index vector. It is acquired and supplied to the scale calculation means 104 (step S102).
- the scale calculation means 104 calculates the similarity ⁇ B between the feature amounts corresponding to the respective quantization indexes from the first quantization index value and the second quantization index value (step S104). .
- the representative value of the feature value before quantization may be calculated from the quantization index, and a value that increases as the difference between the representative values decreases may be used as ⁇ B.
- a table that can subtract the value of ⁇ B by the combination of the quantization index values is held, and this table is obtained from the combination of the quantization index values.
- the value of ⁇ B may be directly obtained by using it.
- step S106 the value of ⁇ B is added to the variable B (step S106).
- the value of ⁇ B is 0, instead of adding 0 to the variable B, it may be controlled not to add.
- step S108 it is checked whether or not the dimension number n has reached the number of dimensions N (step S108). If not, the process proceeds to step S112. If the dimension number n has been reached, the value of the variable B at that time is determined as an identity measure. (Scale representing similarity) is output (step S110), and the process is terminated.
- extraction processing is performed in order from dimensions 1 to N, but the order is not limited to this and may be arbitrary.
- FIG. 23 is another flowchart showing the operation of the verification unit 100. Also in the flowchart of FIG. 23, it is assumed that the dimension (number) of the quantization index vector is represented by n, and there are a total of N dimensions from 1 to N. A variable for calculating the distance scale is represented by C.
- steps S100, S104, S106, and S110 are replaced with steps S200, S204, S206, and S210, respectively.
- variable C is set to zero.
- the representative value of the feature value before quantization may be calculated from the quantization index, and a value that becomes smaller as the difference between the representative values is smaller may be used as ⁇ C.
- a table that can subtract the value of ⁇ C by the combination of the quantization index values is held, and this table is obtained from the combination of the quantization index values. The value of ⁇ C may be directly obtained by using it.
- step S206 the value of ⁇ C is added to the variable C.
- the value of ⁇ C is 0, instead of adding 0 to the variable C, it may be controlled not to add.
- step S210 the value of the variable C at that time is output as an identity scale (a scale representing distance), and the process ends.
- step S108 the process proceeds to step S210.
- extraction processing is performed in order from dimensions 1 to N, but the order is not limited to this and may be arbitrary.
- FIG. 24 is another flowchart showing the operation of the verification unit 100. Also in the flowchart of FIG. 24, the dimension (number) of the quantization index vector is represented by n, and there are total N dimensions from 1 to N. In addition, a variable for calculating the similarity is represented by B, and a variable for counting dimensions where “both quantization indexes are not 0” is represented by A.
- a and B are set to 0 (step S300), and the process proceeds to step S102.
- Step S102 is the same as that in FIG. 22, and after completion, the process proceeds to step S314.
- step S314 the scale calculation means 104 checks whether both the first quantization index value and the second quantization index value are zero. When both are 0, the process proceeds to step S108, and when either one is not 0, the value of the variable A is increased by 1 (step S316), and the process proceeds to step S104.
- steps S104, S106, S108, and S112 are the same as those in FIG. If the dimension number n reaches the dimension number N in step S108, the process proceeds to step S310.
- the identity scale may be calculated as (B / f (D)) + g (D).
- FIG. 26 shows the flow in this case. 24 is basically the same as FIG. 24, except that variable D is set to 0 instead of A in step S500, and if both quantization indexes are 0 in step S314, D is incremented by 1 in step S516, and step S108.
- the process proceeds to step S104, and the difference is that the identity measure is calculated from B and D by the above method in step S510.
- FIG. 25 is another flowchart showing the operation of the verification unit 100. Also in the flowchart of FIG. 25, it is assumed that the dimension (number) of the quantization index vector is represented by n, and there are a total of N dimensions from 1 to N. In addition, a variable for calculating the distance measure is represented by C, and a variable for counting dimensions where “both quantization indexes are not 0” is represented by A.
- steps S300, S104, S106, and S310 are replaced with steps S400, S204, S206, and S410, respectively.
- variable A and variable C are set to zero.
- Step S204 and step S206 are the same as in FIG.
- step S108 the process proceeds to step S410.
- extraction processing is performed in order from dimensions 1 to N, but the order is not limited to this and may be arbitrary.
- D NA may be obtained, and the identity scale may be calculated by C / f (D) or the like.
- the value of D may be directly calculated without obtaining A.
- the present invention relates to a patent application of Japanese Patent Application No. 2009-61021 filed on March 13, 2009 in Japan, and Japanese Patent Application No. 2009- filed on April 14, 2009 in Japan.
- the benefit of the priority claim based on the 97863 patent application is enjoyed, and all the contents described in the patent application are included in this specification.
Abstract
Description
特許文献1に記載されているような、画像の複数の局所領域から抽出した特徴量から成る特徴ベクトルで構成されている画像識別子は、画像の局所領域間の相関が大きい画像に対して、各次元において特徴量を抽出する局所領域の形状が同一であるため(特許文献1の例では同一の形状の長方形領域)、抽出される特徴量の次元間の相関が大きくなる。そのため、画像識別子(特徴ベクトル)の識別能力が低くなる、という第1の問題点がある。ここで形状が同一とは、領域の大きさや角度(傾き或いは姿勢)も含めて同一であるということである。
特許文献1に記載されている画像識別子の第2の問題点は、特徴量(特徴ベクトル)を算出するための各次元の領域の形状(大きさ、角度も含めて)が同一の長方形であるため、長方形の辺の長さと同じ、あるいは、その整数分の1の周期を持つ周波数成分を検知できないという、周波数上の盲点が存在するということである。その理由は、この特定の周波数の信号成分について領域内で平均をとると、信号成分の大小によらず0となってしまい、その周波数成分の信号を全く検知できなくなるためである。より具体的には、長方形の辺の長さと同じ周期を持つ周波数をf0とすると,周波数nf0(n=1,2,3,…)の成分が検知できなくなる。このため、直流成分とこの周波数成分に信号が集中している画像に対しては、画素値の平均値は直流成分と同じになってしまい、領域間で値の差がなくなる結果、領域間の平均画素値の差として抽出される特徴量は全て0になってしまい、識別できなくなる(識別能力が著しく低下する)。実際には、周波数nf0(n=1,2,3,…)の成分のみではなく、その近傍の一定の周波数領域に対しては同様に検知困難となるため、上記特定周波数に信号成分が集中していなくても、その周波数帯の信号成分が使えないことにより、識別能力が低下する。この問題を軽減するには、周波数f0の値を大きくし、上記検知困難な周波数帯に陥る信号電力を下げることが考えられる。しかしながら、周波数f0の値を大きくすることは、領域の大きさを小さくすることを意味し、特徴量の頑健性(各種改変処理やノイズに対して特徴量が変化しない度合い)の低下につながる。例えば、領域が小さくなることで、多少の位置ずれに対しても、特徴量の値が大きく変化することになり、特徴量の頑健性が下がる。このように、同一の長方形領域を用いる場合には、識別能力をあげた上で頑健性を確保することが極めて難しい。
本発明の目的は、上述した課題、すなわち異なる画像を識別できる度合いである識別能力の低い画像識別子を用いた照合では照合精度が低下する、という課題を解決する画像識別子照合装置を提供することにある。
[第1の実施の形態の構成]
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。
第1および第2の抽出領域の必須条件は、次元間で抽出領域対の相対的な位置が異なることに加えて、次元間で抽出領域対の形状の組み合わせが異なることである。
(1)擬似乱数の種(シード)が次元別抽出領域情報として供給される。
(2)次元n=1とする。
(3)擬似乱数を発生させ、次元nの第1の抽出領域の四角形の四隅を決定する。
(4)擬似乱数を発生させ、次元nの第2の抽出領域の四角形の四隅を決定する。
(5)次元n=n+1として、(3)へ戻る。
Qn=+1 (Vn1>Vn2 の場合)
-1 (Vn1≦Vn2 の場合)
Qn=+1 (|Vn1-Vn2|>th かつ Vn1>Vn2 の場合)
0 (|Vn1-Vn2|≦th の場合)
-1 (|Vn1-Vn2|>th かつ Vn1≦Vn2 の場合)
次に、図5のフローチャートを参照して、第1の実施の形態における画像識別子抽出装置の動作を説明する。図5のフローチャートでは、特徴ベクトルの次元(の番号)をnで表し、次元は1からNまでの合計N次元あるものとする。
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。
[第2の実施の形態の構成]
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。
次に、図7のフローチャートを参照して、第2の実施の形態における画像識別子抽出装置の動作を説明する。図7のフローチャートでは、特徴ベクトルの次元(の番号)をnで表し、次元は1からNまでの合計N次元あるものとする。
第2の実施の形態では、量子化の境界が固定されている第1の実施の形態と比較して、量子化の境界が画像に対して適応的に(動的に)算出される点が異なる。第1の実施の形態のように、量子化の境界が固定化されていると、特定の画像(例えば起伏の少ない平坦な画像など)に対して、特徴ベクトルの次元の値が、特定の量子化インデックスに偏る(特定の量子化インデックスの出現確率が高い)という事態が発生し(エントロピーが低くなる)、これらの画像に対して識別能力が低下するという問題が発生する。一方で第2の実施の形態のように、量子化の境界が画像に対して適応的に(動的に)算出されることにより、どの画像に対しても、特徴ベクトルの次元の値が、特定の量子化インデックスに偏る(特定の量子化インデックスの出現確率が高い)ことを抑えることができるため、識別能力を高くすることができる。
[第3の実施の形態の構成]
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。
次に、図10のフローチャートを参照して、第3の実施の形態における画像識別子抽出装置の動作を説明する。図10のフローチャートでは、特徴ベクトルの次元(の番号)をnで表し、次元は1からNまでの合計N次元あるものとする。
第1の実施の形態の効果に加えて、異なる画像を識別できる度合いである識別能力を更に高くすることができる。
[第4の実施の形態の構成]
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。
次に、図13のフローチャートを参照して、第4の実施の形態における画像識別子抽出装置の動作を説明する。図13のフローチャートでは、特徴ベクトルの次元(の番号)をnで表し、次元は1からNまでの合計N次元あるものとする。
第1の実施の形態の効果に加えて、異なる画像を識別できる度合いである識別能力を更に高くすることができる。
以上本発明の実施の形態について説明したが、本発明は以上の実施の形態にのみ限定されず、その他各種の付加変更が可能である。また、本発明の画像識別子抽出装置は、その有する機能をハードウェア的に実現することは勿論、コンピュータとプログラムとで実現することができる。プログラムは、磁気ディスクや半導体メモリ等のコンピュータ可読記録媒体に記録されて提供され、コンピュータの立ち上げ時などにコンピュータに読み取られ、そのコンピュータの動作を制御することにより、そのコンピュータを前述した各実施の形態における次元決定手段、抽出領域取得手段、領域特徴量算出手段、比較手段、領域特徴量算出方法取得手段、比較方法取得手段として機能させる。
第5の実施の形態では、抽出する特徴ベクトルの次元数は300次元(第1次元から第300次元)である。
第6の実施の形態は、第5の実施の形態と同じく、抽出する特徴ベクトルの次元数は300次元(第1次元から第300次元)である。また第6の実施の形態では、抽出領域取得手段2に入力として供給される次元別抽出領域情報として、第5の実施の形態と同じく図14に示す情報を使用する。さらに第6の実施の形態では、比較方法取得手段6に入力として供給される次元別比較・量子化方法情報として、第5の実施の形態と同じく図17に示す情報を使用する。
第7の実施の形態は、抽出する特徴ベクトルの次元数は325次元(第1次元から第325次元)である。第7の実施の形態の場合は、各領域は、画像を縦方向32、横方向32に分割してできる1024個のブロックの組み合わせによって構成されている。ここで、各ブロックに対して、図27に示すように、左上から順に0から始まるインデックスを付与し、このインデックスを用いて領域を記述する。具体的には、長方形領域を、その左上のブロックのインデックスaと右下のブロックのインデックスbを用いてa-bのように表現する。例えば、インデックス0、1、32、33の4つのブロックからなる長方形は、0-33のように記述する。また、このようにしてできる長方形を記号“|”によって繋げた場合は、その記号の前後の長方形を連結してできる領域を表現するものとする。例えば、0-33|2-67は、0-33で定義される長方形と、2-67で定義される長方形を連結してできる領域、すなわち、ブロック番号0、1、2、3、32、33、34、35、66、67によって構成される領域を表している。
次に、本発明で出力される量子化インデックスベクトルを照合する照合手段についてブロック図を用いて説明する。
2…抽出領域取得手段
3、3A…領域特徴量算出手段
31、31A…第1の領域特徴量算出手段
32、32A…第2の領域特徴量算出手段
4、4B…比較手段
41…大小比較手段
42、44…量子化手段
43…差分値算出手段
45…量子化境界決定手段
5…領域特徴量算出方法取得手段
6…比較方法取得手段
Claims (23)
- 画像中の、複数の部分領域対の、各部分領域から領域特徴量を抽出し、部分領域対ごとに、対をなす2つの部分領域の前記領域特徴量の差分値を量子化し、該量子化の際には差分値の絶対値がある規定値より小さい場合は特定の量子化値に量子化し、部分領域対ごとに算出された量子化値を要素とした集合を前記画像の識別に用いる画像識別子とする、生成方法によって生成された、第1の画像の画像識別子と第2の画像の画像識別子とを、前記特定の量子化値である要素の重みが小さくなる方法で照合する照合手段、
を備えることを特徴とする画像識別子照合装置。 - 前記照合手段は、前記第1の画像の画像識別子と前記第2の画像の画像識別子との対応する要素を比較して、前記第1の画像と前記第2の画像とが同一である度合いを示す同一性尺度を算出する
ことを特徴とする請求項1に記載の画像識別子照合装置。 - 前記算出方法として、双方の量子化値がともに前記特定の量子化値である要素間の比較結果の重みが小さくなる算出方法を使用する
ことを特徴とする請求項1または2に記載の画像識別子照合装置。 - 前記照合手段は、少なくとも一方の量子化値が前記特定の量子化値ではない要素の数をAとし、少なくとも一方の量子化値が前記特定の量子化値ではない要素のうち量子化値が一致する要素の数をB、または一致しない要素の数をCとするとき、B/AまたはC/Aの計算結果から前記同一性尺度を算出する
ことを特徴とする請求項3に記載の画像識別子照合装置。 - 前記照合手段は、双方の量子化値がともに前記特定の量子化値である要素の数をDとし、少なくとも一方の量子化値が前記特定の量子化値ではない要素のうち量子化値が一致する要素の数をB、または一致しない要素の数をCとし、任意のDに対して単調非増加な関数をf(D)とするとき、B/f(D)またはC/f(D)の値から前記同一性尺度を算出する
ことを特徴とする請求項3に記載の画像識別子照合装置。 - 前記照合手段は、双方の量子化値がともに前記特定の量子化値である要素の数をDとし、少なくとも一方の量子化値が前記特定の量子化値ではない要素のうち量子化値が一致する要素の数をBとするとき、Dの重みがBの重みよりも小さい重み付けによってDとBを総和した値から前記同一性尺度を算出する
ことを特徴とする請求項3に記載の画像識別子照合装置。 - 前記算出方法として、少なくとも一方の量子化値が前記特定の量子化値である要素間の比較結果の重みが小さくなる算出方法を使用する
ことを特徴とする請求項1または2に記載の画像識別子照合装置。 - 前記照合手段は、双方の量子化値がともに前記特定の量子化値ではない要素の数をAとし、双方の量子化値がともに前記特定の量子化値ではない要素のうち量子化値が一致する要素の数をB、または一致しない要素の数をCとするとき、B/AまたはC/Aの計算結果から前記同一性尺度を算出する
ことを特徴とする請求項7に記載の画像識別子照合装置。 - 前記照合手段は、少なくとも一方の量子化値が前記特定の量子化値である要素の数をDとし、双方の量子化値がともに前記特定の量子化値ではない要素のうち量子化値が一致する要素の数をB、または一致しない要素の数をCとし、任意のDに対して単調非増加な関数をf(D)とするとき、B/f(D)またはC/f(D)の値から前記同一性尺度を算出する
ことを特徴とする請求項7に記載の画像識別子照合装置。 - 前記照合手段は、少なくとも一方の量子化値が前記特定の量子化値である要素の数をDとし、双方の量子化値がともに前記特定の量子化値ではない要素のうち量子化値が一致する要素の数をBとするとき、Dの重みがBの重みよりも小さい重み付けによってDとBを総和した値から前記同一性尺度を算出する
ことを特徴とする請求項7に記載の画像識別子照合装置。 - 前記複数の部分領域対は、対をなす2つの部分領域の形状の組み合わせと、対をなす2つの部分領域の相対的な位置関係との双方が、他の少なくとも1つの部分領域対と相違する1以上の部分領域対を含む、
ことを特徴とする請求項1乃至10に記載の画像識別子照合装置。 - 画像中の、複数の部分領域対の、各部分領域から領域特徴量を抽出し、部分領域対ごとに、対をなす2つの部分領域の前記領域特徴量の差分値を量子化し、該量子化の際には差分値の絶対値がある規定値より小さい場合は特定の量子化値に量子化し、部分領域対ごとに算出された量子化値を要素とした集合を前記画像の識別に用いる画像識別子とする、生成方法によって生成された、第1の画像の画像識別子と第2の画像の画像識別子とを、前記特定の量子化値である要素の重みが小さくなる方法で照合する、
ことを特徴とする画像識別子照合方法。 - 前記第1の画像の画像識別子と前記第2の画像の画像識別子との対応する要素を比較して、前記第1の画像と前記第2の画像とが同一である度合いを示す同一性尺度を算出する
ことを特徴とする請求項12に記載の画像識別子照合方法。 - 前記算出方法として、双方の量子化値がともに前記特定の量子化値である要素間の比較結果の重みが小さくなる算出方法を使用する
ことを特徴とする請求項12または13に記載の画像識別子照合方法。 - 少なくとも一方の量子化値が前記特定の量子化値ではない要素の数をAとし、少なくとも一方の量子化値が前記特定の量子化値ではない要素のうち量子化値が一致する要素の数をB、または一致しない要素の数をCとするとき、B/AまたはC/Aの計算結果から前記同一性尺度を算出する
ことを特徴とする請求項14に記載の画像識別子照合方法。 - 双方の量子化値がともに前記特定の量子化値である要素の数をDとし、少なくとも一方の量子化値が前記特定の量子化値ではない要素のうち量子化値が一致する要素の数をB、または一致しない要素の数をCとし、任意のDに対して単調非増加な関数をf(D)とするとき、B/f(D)またはC/f(D)の値から前記同一性尺度を算出する
ことを特徴とする請求項14に記載の画像識別子照合方法。 - 双方の量子化値がともに前記特定の量子化値である要素の数をDとし、少なくとも一方の量子化値が前記特定の量子化値ではない要素のうち量子化値が一致する要素の数をBとするとき、Dの重みがBの重みよりも小さい重み付けによってDとBを総和した値から前記同一性尺度を算出する
ことを特徴とする請求項14に記載の画像識別子照合方法。 - 前記算出方法として、少なくとも一方の量子化値が前記特定の量子化値である要素間の比較結果の重みが小さくなる算出方法を使用する
ことを特徴とする請求項12または13に記載の画像識別子照合方法。 - 双方の量子化値がともに前記特定の量子化値ではない要素の数をAとし、双方の量子化値がともに前記特定の量子化値ではない要素のうち量子化値が一致する要素の数をB、または一致しない要素の数をCとするとき、B/AまたはC/Aの計算結果から前記同一性尺度を算出する
ことを特徴とする請求項18に記載の画像識別子照合方法。 - 少なくとも一方の量子化値が前記特定の量子化値である要素の数をDとし、双方の量子化値がともに前記特定の量子化値ではない要素のうち量子化値が一致する要素の数をB、または一致しない要素の数をCとし、任意のDに対して単調非増加な関数をf(D)とするとき、B/f(D)またはC/f(D)の値から前記同一性尺度を算出する
ことを特徴とする請求項18に記載の画像識別子照合方法。 - 少なくとも一方の量子化値が前記特定の量子化値である要素の数をDとし、双方の量子化値がともに前記特定の量子化値ではない要素のうち量子化値が一致する要素の数をBとするとき、Dの重みがBの重みよりも小さい重み付けによってDとBを総和した値から前記同一性尺度を算出する
ことを特徴とする請求項18に記載の画像識別子照合方法。 - 前記複数の部分領域対は、対をなす2つの部分領域の形状の組み合わせと、対をなす2つの部分領域の相対的な位置関係との双方が、他の少なくとも1つの部分領域対と相違する1以上の部分領域対を含む、
ことを特徴とする請求項12乃至21に記載の画像識別子照合方法。 - コンピュータを、
画像中の、複数の部分領域対の、各部分領域から領域特徴量を抽出し、部分領域対ごとに、対をなす2つの部分領域の前記領域特徴量の差分値を量子化し、該量子化の際には差分値の絶対値がある規定値より小さい場合は特定の量子化値に量子化し、部分領域対ごとに算出された量子化値を要素とした集合を前記画像の識別に用いる画像識別子とする、生成方法によって生成された、第1の画像の画像識別子と第2の画像の画像識別子とを、前記特定の量子化値である要素の重みが小さくなる方法で照合する照合手段、
として機能させるためのプログラム。
Priority Applications (5)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
EP10750604.0A EP2360639B1 (en) | 2009-03-13 | 2010-03-12 | Image identifier matching device |
US13/131,798 US8270724B2 (en) | 2009-03-13 | 2010-03-12 | Image signature matching device |
KR1020117021098A KR101398272B1 (ko) | 2009-03-13 | 2010-03-12 | 화상 식별자 조합 장치 |
CN201080011021.3A CN102349091B (zh) | 2009-03-13 | 2010-03-12 | 图像签名匹配设备 |
JP2011503731A JP4735783B2 (ja) | 2009-03-13 | 2010-03-12 | 画像識別子照合装置 |
Applications Claiming Priority (4)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
JP2009061021 | 2009-03-13 | ||
JP2009-061021 | 2009-03-13 | ||
JP2009097863 | 2009-04-14 | ||
JP2009-097863 | 2009-04-14 |
Publications (1)
Publication Number | Publication Date |
---|---|
WO2010103848A1 true WO2010103848A1 (ja) | 2010-09-16 |
Family
ID=42728145
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/JP2010/001768 WO2010103848A1 (ja) | 2009-03-13 | 2010-03-12 | 画像識別子照合装置 |
Country Status (6)
Country | Link |
---|---|
US (1) | US8270724B2 (ja) |
EP (1) | EP2360639B1 (ja) |
JP (1) | JP4735783B2 (ja) |
KR (1) | KR101398272B1 (ja) |
CN (1) | CN102349091B (ja) |
WO (1) | WO2010103848A1 (ja) |
Cited By (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JP2013218530A (ja) * | 2012-04-09 | 2013-10-24 | Morpho Inc | 特徴点検出装置、特徴点検出方法、特徴点検出プログラム及び記録媒体 |
US8863165B2 (en) | 2010-11-01 | 2014-10-14 | Gracenote, Inc. | Method and system for presenting additional content at a media system |
US9015741B2 (en) | 2009-04-17 | 2015-04-21 | Gracenote, Inc. | Method and system for remotely controlling consumer electronic devices |
Families Citing this family (32)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US8744193B2 (en) * | 2009-03-13 | 2014-06-03 | Nec Corporation | Image signature extraction device |
US8526743B1 (en) | 2010-11-01 | 2013-09-03 | Raf Technology, Inc. | Defined data patterns for object handling |
US9152862B2 (en) * | 2011-09-15 | 2015-10-06 | Raf Technology, Inc. | Object identification and inventory management |
US9443298B2 (en) | 2012-03-02 | 2016-09-13 | Authentect, Inc. | Digital fingerprinting object authentication and anti-counterfeiting system |
US8774455B2 (en) | 2011-03-02 | 2014-07-08 | Raf Technology, Inc. | Document fingerprinting |
US9870517B2 (en) * | 2011-12-08 | 2018-01-16 | Excalibur Ip, Llc | Image object retrieval |
US9418314B2 (en) * | 2012-01-30 | 2016-08-16 | Nec Corporation | Information processing apparatus and control method and control program thereof, and communication terminal and control method and control program thereof |
US9041727B2 (en) | 2012-03-06 | 2015-05-26 | Apple Inc. | User interface tools for selectively applying effects to image |
US9131192B2 (en) | 2012-03-06 | 2015-09-08 | Apple Inc. | Unified slider control for modifying multiple image properties |
US10282055B2 (en) | 2012-03-06 | 2019-05-07 | Apple Inc. | Ordered processing of edits for a media editing application |
CN103377473B (zh) * | 2012-04-19 | 2017-10-24 | 深圳市世纪光速信息技术有限公司 | 一种图像排重方法和装置 |
US9501684B1 (en) | 2013-12-13 | 2016-11-22 | Google Inc. | Providing non-destructive editing reconciliation |
US10621594B2 (en) | 2016-02-19 | 2020-04-14 | Alitheon, Inc. | Multi-level authentication |
EP3236401A1 (en) | 2016-04-18 | 2017-10-25 | Alitheon, Inc. | Authentication-triggered processes |
US10614302B2 (en) | 2016-05-26 | 2020-04-07 | Alitheon, Inc. | Controlled authentication of physical objects |
US10740767B2 (en) | 2016-06-28 | 2020-08-11 | Alitheon, Inc. | Centralized databases storing digital fingerprints of objects for collaborative authentication |
US10915612B2 (en) | 2016-07-05 | 2021-02-09 | Alitheon, Inc. | Authenticated production |
US10902540B2 (en) | 2016-08-12 | 2021-01-26 | Alitheon, Inc. | Event-driven authentication of physical objects |
US10839528B2 (en) | 2016-08-19 | 2020-11-17 | Alitheon, Inc. | Authentication-based tracking |
EP3435287A3 (en) | 2017-07-25 | 2019-05-01 | Alitheon, Inc. | Model-based digital fingerprinting |
EP3514715A1 (en) | 2018-01-22 | 2019-07-24 | Alitheon, Inc. | Secure digital fingerprint key object database |
US10963670B2 (en) | 2019-02-06 | 2021-03-30 | Alitheon, Inc. | Object change detection and measurement using digital fingerprints |
EP3734506A1 (en) | 2019-05-02 | 2020-11-04 | Alitheon, Inc. | Automated authentication region localization and capture |
EP3736717A1 (en) | 2019-05-10 | 2020-11-11 | Alitheon, Inc. | Loop chain digital fingerprint method and system |
CN110245615B (zh) * | 2019-06-17 | 2022-09-13 | 河海大学常州校区 | 一种签名鉴伪方法、系统及存储介质 |
US11238146B2 (en) | 2019-10-17 | 2022-02-01 | Alitheon, Inc. | Securing composite objects using digital fingerprints |
EP3859603A1 (en) | 2020-01-28 | 2021-08-04 | Alitheon, Inc. | Depth-based digital fingerprinting |
US11568683B2 (en) | 2020-03-23 | 2023-01-31 | Alitheon, Inc. | Facial biometrics system and method using digital fingerprints |
US11341348B2 (en) | 2020-03-23 | 2022-05-24 | Alitheon, Inc. | Hand biometrics system and method using digital fingerprints |
US11948377B2 (en) | 2020-04-06 | 2024-04-02 | Alitheon, Inc. | Local encoding of intrinsic authentication data |
US11663849B1 (en) | 2020-04-23 | 2023-05-30 | Alitheon, Inc. | Transform pyramiding for fingerprint matching system and method |
US11700123B2 (en) | 2020-06-17 | 2023-07-11 | Alitheon, Inc. | Asset-backed digital security tokens |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JPH08500471A (ja) | 1992-04-30 | 1996-01-16 | セリディアン コーポレイション | 放送セグメントを認識するための方法とシステム |
JP2006268825A (ja) * | 2005-02-28 | 2006-10-05 | Toshiba Corp | オブジェクト検出装置、学習装置、オブジェクト検出システム、方法、およびプログラム |
JP2007140684A (ja) * | 2005-11-15 | 2007-06-07 | Toshiba Corp | 画像処理装置、方法、プログラム |
JP2009061021A (ja) | 2007-09-05 | 2009-03-26 | Univ Kansai | 分岐型生分解性ポリエステルを含む生体内留置物 |
JP2009097863A (ja) | 2007-10-12 | 2009-05-07 | Toshiba Corp | 機器位置検出システム |
Family Cites Families (18)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US4739398A (en) * | 1986-05-02 | 1988-04-19 | Control Data Corporation | Method, apparatus and system for recognizing broadcast segments |
US5157743A (en) * | 1987-10-28 | 1992-10-20 | Canon Kabushiki Kaisha | Image information coding apparatus |
US5067189A (en) * | 1990-04-11 | 1991-11-26 | Weedling Robert E | Air chamber type patient mover air pallet with multiple control features |
US5561873A (en) * | 1994-07-15 | 1996-10-08 | Patient Transfer Systems, Inc. | Air chamber-type patient mover air pallet with multiple control features |
US6073291A (en) * | 1997-02-21 | 2000-06-13 | Davis; David T. | Inflatable medical patient transfer apparatus |
DE69940703D1 (de) * | 1998-03-05 | 2009-05-20 | Panasonic Corp | Verfahren zur Bildkodierung, Verfahren zur Bildkodierung/-Dekodierung, Bildkodierer, oder Vorrichtung zur Bildaufzeichnung/-Wiedergabe |
JP3684148B2 (ja) * | 2000-10-20 | 2005-08-17 | キヤノン株式会社 | 画像処理方法及びその装置、記憶媒体 |
KR100450793B1 (ko) * | 2001-01-20 | 2004-10-01 | 삼성전자주식회사 | 영역 분할된 영상의 영역 특징치 정합에 기초한객체추출장치 및 그 방법 |
JP4190739B2 (ja) * | 2001-02-19 | 2008-12-03 | 日本電気株式会社 | 画像特徴量生成装置および方法、ならびに画像特徴量生成プログラムを記録した記憶媒体 |
JP4161659B2 (ja) * | 2002-02-27 | 2008-10-08 | 日本電気株式会社 | 画像認識システム及びその認識方法並びにプログラム |
US7643684B2 (en) * | 2003-07-15 | 2010-01-05 | Samsung Electronics Co., Ltd. | Apparatus for and method of constructing multi-view face database, and apparatus for and method of generating multi-view face descriptor |
US6898809B2 (en) * | 2003-08-11 | 2005-05-31 | Woodlark Circle, Inc. | Air mattress with single perimeter seam |
US7712170B2 (en) * | 2003-08-11 | 2010-05-11 | Woodlark Circle, Inc. | Single patient, personal use air mattress having a single perimeter seam |
US7379627B2 (en) * | 2003-10-20 | 2008-05-27 | Microsoft Corporation | Integrated solution to digital image similarity searching |
US7986813B2 (en) * | 2004-03-03 | 2011-07-26 | Nec Corporation | Object pose estimation and comparison system using image sharpness differences, object pose estimation and comparison method using image sharpness differences, and program therefor |
US20060193520A1 (en) * | 2005-02-28 | 2006-08-31 | Takeshi Mita | Object detection apparatus, learning apparatus, object detection system, object detection method and object detection program |
JP2007293438A (ja) * | 2006-04-21 | 2007-11-08 | Omron Corp | 特徴量取得装置 |
US7805011B2 (en) * | 2006-09-13 | 2010-09-28 | Warner Bros. Entertainment Inc. | Method and apparatus for providing lossless data compression and editing media content |
-
2010
- 2010-03-12 WO PCT/JP2010/001768 patent/WO2010103848A1/ja active Application Filing
- 2010-03-12 KR KR1020117021098A patent/KR101398272B1/ko active IP Right Grant
- 2010-03-12 JP JP2011503731A patent/JP4735783B2/ja active Active
- 2010-03-12 CN CN201080011021.3A patent/CN102349091B/zh active Active
- 2010-03-12 EP EP10750604.0A patent/EP2360639B1/en active Active
- 2010-03-12 US US13/131,798 patent/US8270724B2/en active Active
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JPH08500471A (ja) | 1992-04-30 | 1996-01-16 | セリディアン コーポレイション | 放送セグメントを認識するための方法とシステム |
JP2006268825A (ja) * | 2005-02-28 | 2006-10-05 | Toshiba Corp | オブジェクト検出装置、学習装置、オブジェクト検出システム、方法、およびプログラム |
JP2007140684A (ja) * | 2005-11-15 | 2007-06-07 | Toshiba Corp | 画像処理装置、方法、プログラム |
JP2009061021A (ja) | 2007-09-05 | 2009-03-26 | Univ Kansai | 分岐型生分解性ポリエステルを含む生体内留置物 |
JP2009097863A (ja) | 2007-10-12 | 2009-05-07 | Toshiba Corp | 機器位置検出システム |
Non-Patent Citations (1)
Title |
---|
See also references of EP2360639A4 |
Cited By (41)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US10972763B2 (en) | 2009-04-17 | 2021-04-06 | Gracenote, Inc. | Method and system for remotely controlling consumer electronic device |
US10341697B2 (en) | 2009-04-17 | 2019-07-02 | Gracenote, Inc. | Method and system for remotely controlling consumer electronic devices |
US9015741B2 (en) | 2009-04-17 | 2015-04-21 | Gracenote, Inc. | Method and system for remotely controlling consumer electronic devices |
US11856155B2 (en) | 2009-04-17 | 2023-12-26 | Roku, Inc. | Method and system for remotely controlling consumer electronic devices |
US9992518B2 (en) | 2009-04-17 | 2018-06-05 | Gracenote, Inc. | Method and system for remotely controlling consumer electronic devices |
US10979742B2 (en) | 2009-04-17 | 2021-04-13 | Gracenote, Inc. | Method and system for remotely controlling consumer electronic device |
US10250919B2 (en) | 2009-04-17 | 2019-04-02 | Gracenote, Inc. | Method and system for remotely controlling consumer electronic devices |
US11818403B2 (en) | 2009-04-17 | 2023-11-14 | Roku, Inc. | Method and system for remotely controlling consumer electronic devices |
US11611783B2 (en) | 2009-04-17 | 2023-03-21 | Roku, Inc. | Method and system for remotely controlling consumer electronic device |
US11297359B2 (en) | 2009-04-17 | 2022-04-05 | Roku, Inc. | Method and system for remotely controlling consumer electronic devices |
US10701411B2 (en) | 2009-04-17 | 2020-06-30 | Gracenote, Inc. | Method and system for remotely controlling consumer electronic devices |
US10701410B2 (en) | 2009-04-17 | 2020-06-30 | Gracenote, Inc. | Method and system for remotely controlling consumer electronic device |
US10701412B2 (en) | 2009-04-17 | 2020-06-30 | Gracenote, Inc. | Method and system for remotely controlling consumer electronic devices |
US10715841B2 (en) | 2009-04-17 | 2020-07-14 | Gracenote, Inc. | Method and system for remotely controlling consumer electronic devices |
US10735782B2 (en) | 2009-04-17 | 2020-08-04 | Gracenote, Inc. | Method and system for remotely controlling consumer electronic devices |
US11206435B2 (en) | 2009-04-17 | 2021-12-21 | Roku, Inc. | Method and system for remotely controlling consumer electronic devices |
US10904589B2 (en) | 2009-04-17 | 2021-01-26 | Gracenote, Inc. | Method and system for remotely controlling consumer electronic devices |
US11166056B2 (en) | 2009-04-17 | 2021-11-02 | Roku, Inc. | Method and system for remotely controlling consumer electronic devices |
US10972764B2 (en) | 2009-04-17 | 2021-04-06 | Gracenote, Inc. | Method and system for remotely controlling consumer electronic devices |
US10972766B2 (en) | 2009-04-17 | 2021-04-06 | Gracenote, Inc. | Method and system for remotely controlling consumer electronic device |
US9706233B2 (en) | 2009-04-17 | 2017-07-11 | Gracenote, Inc. | Method and system for remotely controlling consumer electronic devices |
US11140425B2 (en) | 2009-04-17 | 2021-10-05 | Roku, Inc. | Method and system for remotely controlling consumer electronic devices |
US9998767B2 (en) | 2009-04-17 | 2018-06-12 | Gracenote, Inc. | Method and system for remotely controlling consumer electronic devices |
US11064223B2 (en) | 2009-04-17 | 2021-07-13 | Roku, Inc. | Method and system for remotely controlling consumer electronic devices |
US11064224B2 (en) | 2009-04-17 | 2021-07-13 | Roku, Inc. | Method and system for remotely controlling consumer electronic devices |
US11064225B2 (en) | 2009-04-17 | 2021-07-13 | Roku, Inc. | Method and system for remotely controlling consumer electronic devices |
US11070852B2 (en) | 2009-04-17 | 2021-07-20 | Roku, Inc. | Method and system for remotely controlling consumer electronic devices |
US11134280B2 (en) | 2009-04-17 | 2021-09-28 | Roku, Inc. | Method and system for remotely controlling consumer electronic devices |
US11134281B2 (en) | 2009-04-17 | 2021-09-28 | Roku, Inc. | Method and system for remotely controlling consumer electronic devices |
US11102550B2 (en) | 2010-11-01 | 2021-08-24 | Roku, Inc. | Method and system for presenting additional content at a media system |
US8863165B2 (en) | 2010-11-01 | 2014-10-14 | Gracenote, Inc. | Method and system for presenting additional content at a media system |
US10904629B2 (en) | 2010-11-01 | 2021-01-26 | Gracenote, Inc. | Method and system for presenting additional content at a media system |
US10869095B2 (en) | 2010-11-01 | 2020-12-15 | Gracenote, Inc. | Method and system for presenting additional content at a media system |
US10506291B2 (en) | 2010-11-01 | 2019-12-10 | Gracenote, Inc. | Method and system for presenting additional content at a media system |
US11336964B2 (en) | 2010-11-01 | 2022-05-17 | Roku, Inc. | Method and system for presenting additional content at a media system |
US11445258B2 (en) | 2010-11-01 | 2022-09-13 | Roku, Inc. | Method and system for presenting additional content at a media system |
US10341734B2 (en) | 2010-11-01 | 2019-07-02 | Gracenote, Inc. | Method and system for presenting additional content at a media system |
US11743543B2 (en) | 2010-11-01 | 2023-08-29 | Roku, Inc. | Method and system for presenting additional content at a media system |
US10979772B2 (en) | 2010-11-01 | 2021-04-13 | Gracenote, Inc. | Method and system for presenting additional content at a media system |
US11889155B2 (en) | 2010-11-01 | 2024-01-30 | Roku, Inc. | Method and system for presenting additional content at a media system |
JP2013218530A (ja) * | 2012-04-09 | 2013-10-24 | Morpho Inc | 特徴点検出装置、特徴点検出方法、特徴点検出プログラム及び記録媒体 |
Also Published As
Publication number | Publication date |
---|---|
US8270724B2 (en) | 2012-09-18 |
JP4735783B2 (ja) | 2011-07-27 |
KR20110123769A (ko) | 2011-11-15 |
US20110235920A1 (en) | 2011-09-29 |
CN102349091A (zh) | 2012-02-08 |
KR101398272B1 (ko) | 2014-06-27 |
EP2360639A4 (en) | 2012-05-02 |
CN102349091B (zh) | 2014-08-06 |
EP2360639A1 (en) | 2011-08-24 |
EP2360639B1 (en) | 2016-11-16 |
JPWO2010103848A1 (ja) | 2012-09-13 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
JP4735783B2 (ja) | 画像識別子照合装置 | |
JP5045817B2 (ja) | 画像識別子抽出装置 | |
JP4735786B2 (ja) | 画像識別子照合装置 | |
JP5573862B2 (ja) | 画像識別子抽出装置 | |
JP5686120B2 (ja) | 画像識別子抽出装置 | |
JP5454600B2 (ja) | 画像識別子抽出装置 |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
WWE | Wipo information: entry into national phase |
Ref document number: 201080011021.3 Country of ref document: CN |
|
121 | Ep: the epo has been informed by wipo that ep was designated in this application |
Ref document number: 10750604 Country of ref document: EP Kind code of ref document: A1 |
|
WWE | Wipo information: entry into national phase |
Ref document number: 2011503731 Country of ref document: JP |
|
WWE | Wipo information: entry into national phase |
Ref document number: 13131798 Country of ref document: US |
|
REEP | Request for entry into the european phase |
Ref document number: 2010750604 Country of ref document: EP |
|
WWE | Wipo information: entry into national phase |
Ref document number: 2010750604 Country of ref document: EP |
|
ENP | Entry into the national phase |
Ref document number: 20117021098 Country of ref document: KR Kind code of ref document: A |
|
NENP | Non-entry into the national phase |
Ref country code: DE |