WO2008026414A1 - Procédé de reconnaissance d'image, dispositif de reconnaissance d'image et programme de reconnaissance d'image - Google Patents
Procédé de reconnaissance d'image, dispositif de reconnaissance d'image et programme de reconnaissance d'image Download PDFInfo
- Publication number
- WO2008026414A1 WO2008026414A1 PCT/JP2007/065086 JP2007065086W WO2008026414A1 WO 2008026414 A1 WO2008026414 A1 WO 2008026414A1 JP 2007065086 W JP2007065086 W JP 2007065086W WO 2008026414 A1 WO2008026414 A1 WO 2008026414A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- image
- local
- search
- recognition
- input image
- Prior art date
Links
Classifications
-
- 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
- G06V10/75—Organisation of the matching processes, e.g. simultaneous or sequential comparisons of image or video features; Coarse-fine approaches, e.g. multi-scale approaches; using context analysis; Selection of dictionaries
- G06V10/757—Matching configurations of points or features
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/211—Selection of the most significant subset of features
- G06F18/2115—Selection of the most significant subset of features by evaluating different subsets according to an optimisation criterion, e.g. class separability, forward selection or backward elimination
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2413—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on distances to training or reference patterns
- G06F18/24147—Distances to closest patterns, e.g. nearest neighbour classification
-
- 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
- G06V10/46—Descriptors for shape, contour or point-related descriptors, e.g. scale invariant feature transform [SIFT] or bags of words [BoW]; Salient regional features
- G06V10/462—Salient features, e.g. scale invariant feature transforms [SIFT]
-
- 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
- G06V10/764—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using classification, e.g. of video objects
-
- 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
- G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
- G06V10/771—Feature selection, e.g. selecting representative features from a multi-dimensional feature space
Definitions
- Image recognition method image recognition apparatus, and image recognition program
- the present invention relates to an image recognition method that performs image recognition using a local descriptor that describes local features of an image, an image recognition apparatus that performs image recognition using a local descriptor, and an image recognition program.
- the target force is not a 3 ⁇ 4-dimensional object but a pattern on the plane (planar object), and an instance (car) that does not recognize the object class (for example, whether the object in the photograph belongs to the category of car).
- the service is already possible. For example, a service by Dai Nippon Printing Co., Ltd. that uses the technology of Clementec Co., Ltd. (US.
- Patent ⁇ ⁇ 20040208372 a service of Olympus Co., Ltd., and NEC Corporation that uses the technology of Evolution Robotics, Inc. Services are known. If such a planar object can be recognized, it will open the way to automatic indexing of existing images and videos, rather than just taking pictures of posters and products.
- a local descriptor describes the image by capturing local features of the image and extracting them as multidimensional feature vectors. Since the value is determined locally, it has the property of being relatively strong (robust) against hiding and image fluctuations.
- “local” means a part of an image. It means that there is a “local descriptor” that represents a partial feature of an image. In this specification, a local descriptor is also called a feature vector.
- the basic operation is to measure the distance between feature vectors obtained from two images and associate them with the nearest one.
- a feature vector is then associated between the image obtained by the camera and a number of images in the database, and the images in the database are voted on. Finally, the label of the image with the largest number of votes is output as the “recognition result”.
- the number of dimensions of the feature vector is on the order of several tens to several hundreds and several hundreds to several thousand per several power images, it is practical to simply calculate the distances of all combinations. I understand that it is not.
- Non-Patent Documents 1 and 2 Recently, recent development of nearest neighbor search technology has made it possible to search a huge number of feature vectors in a short time (for example, see Non-Patent Documents 1 and 2).
- ANN Appro ximate Nearest Neighbor
- LSH Licality Sensitive Hashing
- a high-speed search is realized.
- SR-Tree for accurate nearest neighbor search
- Non-Patent Document 5 there is distributed coding by Kobayashi et al. (For example, see Non-Patent Document 6) as a method of approximate nearest neighbor search.
- Patent Document 1 International Publication No. 2006/092957 Pamphlet
- Non-Patent Document 1 P.indyk, Nearest neighbors in hign-dimensional spaces, Handbook of discrete and computational geometry (Bds. By JE Goodman and J.O'Rourke), Chap man & Hall / CRC, pp.877-892, 2004.
- Patent Document 2 Lr. Shakhnarovich, T. D rrell and P. Indyk Eds Nearest-neighbor meth ods in learning and vision, The MIT Press, 2005.
- Patent Document 3 S. Arya, DM Mount, R. Silverman and AY Wu, ⁇ An optimal algorithm for approximate nearest neighbor searching, Journal of the ACM, vol. 45, no. 6, pp. 891-923, 1998.
- Patent Document 4 M. Datar, N. Immorlica, P. Indyk and VS Mirrokni, Locality-sensitive hasning scheme based on p-staole distributions, Proc. Of the 20th annual symposium on Computational Geometry, pp.253-262, 2004 .
- Non-Patent Document 5 Norio Katayama, Shinichi Sato, Indexing technique for similarity search, "Information processing, vol. 42, no. 10, pp.958-964, Oct., 2001.
- Non-Patent Document 6 Takuo Kobayashi and Masaki Nakagawa, “High-dimensional nearest neighbor search by distributed coding,” IEICE Technical Report PRMU2006-41, June, 2006.
- Non-patent document 7 Toshikazu Wada, learning using non-linear partitioning and learning of non-linear mapping (1) Accelerating nearest neighbor identification using spatial partitioning, "Information processing, vol.46, no.8, pp.912-918 , Aug., 2005.
- Non-patent literature 8 Tomoyuki Shibata, Takekazu Kato, Toshikazu Wada, "Kd decision tree and its application-speeding up nearest neighbor classifier and saving memory," Science theory (D-II) , V01.J88-D-II, no.8, pp.1367-1 377, Aug., 2005.
- the present invention provides
- a search unit for searching for a local descriptor in the input image that is close to each local descriptor in the input image, and specifying a local descriptor in the vicinity for each local descriptor in the input image, and each local in the vicinity The image from which the descriptor was obtained Among them, an identification unit that identifies an image to be a recognition result using a statistical process is provided, and the limiting unit limits the search target to a number that can identify an image to be a recognition result.
- An image recognition apparatus is provided.
- a program that uses a computer V to realize a function for identifying an image including the object from an image database by searching for a local descriptor.
- a feature deriving unit for deriving a plurality of local descriptors representing local features from the input image, and among the local descriptors obtained from the images in the image database, The search is limited to each target to be searched, and the search target that is close to each local descriptor of the input image is searched, and each local descriptor in the vicinity of each local descriptor of the input image is searched. Search unit to identify
- the computer should function as an identification unit that identifies, using statistical processing, an image that should be recognized as a recognition result among images obtained from each local descriptor in the vicinity, and the limiting unit should be a recognition result.
- An image recognition program characterized in that the search target is limited to a number that can identify an image.
- a method for recognizing an image including an object based on a pattern of the object included in an input image from an image database organized using a hash table, wherein the local feature of the pattern A step of extracting one or more feature vectors representing a quantity, an index calculation step of calculating an index of a hash table from the extracted feature vector, and a candidate in the image database by referring to the hash table with the calculated index
- a voting step of deciding an image and voting on the determined candidate image and a step of obtaining an image of a recognition result based on a voting result for each feature vector
- the step of creating the hash table includes: For each feature vector extracted from each image registered in the database, a hash table index is calculated and identified from the feature vectors.
- an image recognition method characterized by including steps of removing feature vectors having low ability and registering image reference data corresponding to each remaining feature vector.
- the present invention (5) An apparatus for recognizing an image including the object based on a pattern of the object included in the input image from an image database organized using a hash table, wherein the local feature of the pattern A feature point extracting unit for extracting one or more feature vectors representing an index, an index calculating unit for calculating an index of a hash table from the extracted feature vector, and referring to the hash table with the calculated index in the image database A voting unit for determining candidate images and voting on the determined candidate images; and an image selection unit for obtaining images of recognition results based on the voting results for each feature vector.
- an index of the hash table is taken into consideration for the variation of the feature value.
- the image is characterized by including the steps of registering image reference data corresponding to each of the remaining feature vectors by removing the feature vectors from each feature vector and removing the feature vectors.
- the limiting step limits the search target to a number that can identify an image to be a recognition result. Processing time required can be reduced. That is, an object can be recognized at high speed.
- the limiting unit limits the search target to a number that can identify an image to be a recognition result.
- the processing time required for recognition can be shortened.
- the limiting unit limits the search target to a number that can identify an image that should be a recognition result, and thus requires recognition. Processing time can be shortened.
- the image recognition method of (4) according to the present invention the low discrimination ability! / And the feature vector are excluded, and only the image reference data corresponding to the feature vector having the high discrimination ability is a hash table. Therefore, it is possible to perform image recognition in a short time using only feature vectors with high discrimination ability as processing targets.
- the memory capacity required for the image database compared to registering image reference data corresponding to all feature vectors. The Save with power S.
- the image recognition device of (5) since only the image reference data corresponding to the feature vector having a high identification capability is registered in the hash table, these are set as processing targets. Image recognition can be performed in a short time. In addition, since only the image reference data corresponding to the feature vector having high identification ability is registered in the hash table, the memory capacity of the image database can be saved.
- the image recognition method of the present invention recognizes an image using a feature vector.
- the basis of recognition is to collate the feature vectors registered in the database with the feature vectors of the input image. Since a feature vector represents a local feature of an image, a plurality of feature vectors are generally obtained from one image.
- some of the feature vectors of objects (images) registered in the force database are those that express the characteristics of the object well (highly discriminating ability) and those that do not (lowly discriminating ability).
- a well-characterized feature of an object is a feature vector that provides sufficient evidence that the input image can be said to be that object if the feature vector is present.
- Feature vector exclusion refers to the latter, that is, the process of deleting feature vectors that cannot be evidence from the dictionary. More specifically,
- the limiting step may be capable of varying a degree of limiting a search target according to an input image so that an image to be a recognition result is identified. . That is, the degree of approximation may be varied depending on the input image. In this way, the power S can be reduced by reducing the processing time by changing the degree of approximation used for recognition according to the image.
- the degree of approximation balances recognition rate and efficiency. It is an important parameter for The stronger the approximation, the more processing time can be reduced. If the approximation is too strong, the nearest neighbors cannot be obtained for many feature vectors, resulting in false recognition.
- One problem here is that the degree of approximation that causes misrecognition varies from image to image. While there are “simple” images that can be recognized even after a large approximation, there are also “difficult” images that are misrecognized. In order to secure a certain recognition rate by fixed approximation, it is difficult to recognize the degree of approximation! / And it is necessary to match the image, which hinders efficiency improvement.
- a method of reducing processing is provided from the viewpoint that "the accuracy of nearest neighbor search necessary for recognition varies depending on images".
- it is a technique that adaptively adjusts the degree of approximation to the image.
- a plurality of discriminators having different degrees of approximation are prepared, and they are connected in cascade in a multi-stage manner from a high degree of approximation to one to a weak one.
- an easily recognizable image can be recognized at a high speed by a large approximate classifier in the former part, and it cannot be recognized! /, Only the image is approximated in the latter part. This is because it is necessary to spend time and sensible words with a discriminator.
- the limiting step relaxes the degree of limiting the search target and determines a new search target by excluding the search target first.
- the search process and the identification process may be executed for the determined search target. In this way, even if the limiting process, search process, and identification process are executed in multiple stages by changing the degree of approximation, the search target in each stage is compared with the case of searching at once. Recognize in a processing time that is not inferior.
- a feature of this method is a method for constructing a multi-stage discriminator.
- the latter classifier only the difference due to the difference in approximation, that is, only the feature vector that was not targeted by the preceding classifier is used as the target of distance calculation, so that the last stage The ability to obtain the advantage of requiring almost the same amount of computation as when using a single classifier.
- the degree of limiting the search target is relaxed in stages, and the limiting process, the searching process, and the search process. If the image that should be the recognition result cannot be identified even after repeating the identification process, the search result for the local descriptor may be rejected. In this way, the misrecognition rate can be suppressed compared to the case where no rejection is performed.
- the image database includes a hash table in which each local descriptor derived from each image is classified by an index value calculated in a predetermined procedure, and the limiting step includes a feature amount Taking the fluctuation into account, the indentation value is calculated from each local descriptor of the input image by the above procedure, the local index belonging to the class is searched by referring to the hash table with the calculated index value, and the identification For each local descriptor specified in the search process, the process uses a statistical process for voting on the image from which it was obtained, and the hash table shows, for each class, the local descriptor belonging to that class. It may be created by excluding the search target power from the local descriptor of that kind when the number of data exceeds the threshold. In this way, if the number of local descriptors belonging to each class exceeds the threshold value, they are excluded from the search target, and the hash table is created. Descriptors are limited to those with high identification ability, enabling efficient recognition
- each local descriptor is represented as a vector
- the process of calculating the index value of the hash table in consideration of the variation of the feature amount is to set the error range to the discrete value obtained by discretizing the elements of each local descriptor.
- the range of the error may be determined according to the variation. In other words, when calculating the index, if the value of the element, the estimated force of fluctuation, and the range of the calculated value span multiple sections used for discretization, the discrete values corresponding to each section are used, Multiple indexes Do it like that.
- the pattern of the object in the image database is a pattern in which the object is viewed from an angle different from that of the input image, that is, if there is a change, the pattern between the image to be recognized and the input image
- the value of the element of the local descriptor (feature vector) in the correspondence relationship changes.
- the value of the local descriptor element is calculated according to a predetermined procedure, and the index value, which is the variance value, is calculated according to a predetermined procedure. If the element value of the feature vector fluctuates, a different discrete value is calculated. It can be said that there is a high possibility of being done.
- Each element of the feature vector is a discrete value discretized with a predetermined threshold.
- the search step calculates a distance between each local descriptor of the input image and each local descriptor in the hash table belonging to the class corresponding to the local descriptor within the predetermined distance or the shortest distance. It may be a process to identify.
- the searching step may be a step in which each local descriptor in the hash table belonging to a class corresponding to each local descriptor of the input image is set as a local descriptor in the vicinity.
- the processing time required for the search can be shortened compared to the case of performing the distance calculation.
- the limiting unit varies the degree of limiting the search target according to the input image so that an image to be a recognition result is identified. You may get it. That is, the degree of approximation may be varied depending on the input image. In this way, it is possible to reduce the processing time by changing the degree of approximation used for recognition according to the image.
- the limiting unit relaxes the degree of limiting the search target, and excludes the search target that has been previously set as a new search target.
- the search unit further performs a process for determining the The local descriptor may be specified, and the identification unit may further identify an image to be a recognition result based on each specified local descriptor. In this way, even when the limiting unit, the search unit, and the identification unit execute multi-step processing by changing the degree of approximation, the search target at each step is searched at once. Compared to the processing time comparable to that of recognition.
- the image is included in the image in the image database.
- the pattern of the object to be displayed may be a pattern when the object is viewed from an angle different from that of the input image.
- the image is included in the image in the image database.
- a part of the pattern of the object to be processed may correspond to the pattern of the input image.
- the identification ability is low! /
- the feature vector is excluded when the feature vector having the same index exceeds a predetermined number is calculated by the index. It may be a process of excluding each feature vector to be registered in the hash table. If there are a large number of feature vectors to be registered for one index of the hash table, it can be said that these feature vectors have low identification ability.
- an index is calculated from the feature vector extracted from the input image and the hash table is referred to by the calculated index, a large number of images that can be candidates are registered for that index! / It is.
- Such an index does not contribute much to narrowing down the images to be recognized. Therefore, for each index, by removing those feature vectors from the registration target in the hash table, only the image reference data with high discriminating power can be registered in the hash table.
- the index calculation step calculates an index by including a discrete value satisfying a range of error estimation in a discrete value obtained by discretizing elements of each feature vector. You may make it take out. In other words, when calculating the index, if the value of the element and the estimated value of fluctuation, and the range of the calculated value spans multiple sections used for discretization, a plurality of discrete values corresponding to each section are used. An index may be calculated.
- the pattern of the object in the image database is a pattern in which the object is viewed from an angle different from that of the input image (there is variation)
- a feature vector having a correspondence relationship between the image to be recognized and the input image The value of the element of changes.
- the element values of the feature vector are discretized based on the threshold value.
- the value of the element of the feature vector in the correspondence relation is near the threshold value, the value is changed if the value varies.
- the possibility of discretization into different discrete values is high. Therefore, when the fluctuation estimation section centering on the value of the element of the feature vector extends over a plurality of sections used for discretization, the above calculation is performed by calculating a plurality of indexes using the discrete values corresponding to each section. A reduction in the recognition rate with respect to fluctuations can be suppressed. In other words, if an element of a feature vector is close to the discretization threshold when calculating the index, the recognition rate can be ensured by calculating the index considering the possibility of crossing the threshold. .
- the image reference data registered in the hash table includes an image ID for identifying an image in the database including each feature vector and elements of the feature vector, and the voting process includes an input image.
- the distance between each feature vector and the feature vector registered in the hash table index calculated from the feature vector is calculated! /, And the image identified by the image ID of the feature vector with the shortest distance is used. It may be a process of voting. In this case, the voting process may be performed only when the shortest distance is equal to or smaller than a certain threshold value.
- the image reference data registered in the hash table includes an image ID for identifying an image in the database including each feature vector, and the voting process is calculated from each feature vector of the input image. It may be a process of voting for an image identified by an image ID registered in the index of the hash table. In this way, only the image ID is registered in the hash table, and it is not necessary to register the elements of each feature vector. Can save even more memory.
- the hash table is referenced with the calculated index, and the recognition process is performed by a simple process that performs a voting process using the image ID registered in the index. Computation time can be further shortened compared to the case of performing.
- the feature vectors with low identification ability are excluded when the feature vectors with the same index exceed a predetermined number, May be excluded from being registered in the hash table.
- the index calculation unit may calculate the index by including a discrete value satisfying a range of error estimation in a scatter value obtained by discretizing elements of each feature vector.
- a plurality of indexes may be calculated using discrete values corresponding to each section.
- the image reference data registered in the hash table includes an image ID for identifying an image in the database including each feature vector and an element of the feature vector. Calculate the distance between each feature vector and each feature vector registered in the hash table index calculated from the feature vector! /, Vote for the image identified by the image ID of the feature vector with the shortest distance May be. In this case, voting may be performed only when the shortest distance is not more than a certain threshold.
- the image reference data registered in the hash table includes an image ID for identifying an image in a database including each feature vector, and the voting unit is calculated from each feature vector of the input image. You may vote for the image identified by the image ID registered in the index of the hash table.
- FIG. 1 is a graph showing a value distribution of a feature vector obtained by a conventional PCA-SIFT.
- FIG. 2 is an explanatory diagram showing the concept of a prior art approximate nearest neighbor search by ANN.
- FIG. 3 is an explanatory diagram showing processing when a collision occurs during data registration according to the present invention!
- FIG. 5 An explanatory diagram showing an example of a search question image used in the experiment according to the present invention. 6] This is a graph showing the results of experiments on the recognition rate and processing time when the tolerance ⁇ was changed from 2 to 100 using the ANN of the prior art.
- FIG. 12 is a graph in which various parameters are changed, the recognition rate is plotted on the horizontal axis, and the processing time is plotted on the vertical axis in order to compare the features of the methods of the present invention and the conventional methods.
- FIG. 13 is a graph showing the relationship between the shooting angle and the recognition rate in each method of the present invention and in the conventional method.
- FIG. 15 is a block diagram showing a configuration example corresponding to a method without distance calculation in the image recognition apparatus of the present invention.
- This block is a block diagram showing a configuration example corresponding to the method with distance calculation in the image recognition apparatus of the present invention.
- FIG. 17 is a diagram showing a discretization method for dealing with fluctuations in the value of each dimension of a feature vector in the present invention.
- FIG. 18 is a block diagram showing a configuration in which discriminators are cascade-connected in multiple stages as one embodiment of the present invention.
- FIG. 19 is a graph showing the relationship between the accuracy of approximate nearest neighbor search and the image recognition rate for the conventional method.
- FIG. 20 is a graph showing the relationship between b, recognition rate, and processing time in the method with distance calculation according to the present invention.
- FIG. 21 is a graph showing the relationship between b, the recognition rate, and the processing time in the method without distance calculation according to the present invention.
- FIG. 22 is a graph showing the relationship between the recognition rate and the processing time for each method in order to compare the characteristics of the method according to the present invention and the conventional method in the case of rejecting! /.
- FIG. 23 is a block diagram showing an image recognition apparatus according to the present invention in which classifiers with distance calculation are connected in cascade in multiple stages.
- FIG. 24 is a block diagram showing an image recognition apparatus according to the present invention in which classifiers without distance calculation are connected in cascade in multiple stages.
- Multi-stage method will be described.
- the process of identifying an image corresponding to the input image is performed by a classifier.
- the discriminator performs processing corresponding to the limiting step, the searching step, and the discriminating step in the method claims.
- it is a part corresponding to a limiting part, a searching part, and an identifying part in the claim of the product and the claim of the program.
- the performance limit is the recognition rate when no approximation is performed.
- the ability to increase the speed with approximation can generally be realized.
- the relationship between the degree of approximation and the recognition rate differs depending on the image to be recognized. Therefore, to reduce the processing time while maintaining the recognition rate, the degree of approximation is adaptively applied. there is a need force s to adjust.
- FIG. 18 is a block diagram showing a configuration in which discriminators are connected in cascade in multiple stages.
- a rectangle with numbers 1 to N represents a discriminator, and the smaller the number, the stronger the approximation.
- Input image power of search query The set of feature vectors obtained is first recognized by the first classifier. If sufficient evidence is obtained at this stage, the recognition process is terminated by the reliability judgment unit and the result is returned. On the other hand, if sufficient evidence cannot be obtained, the set of feature vectors is recognized again by the weaker classifier of the next stage. If enough evidence is not obtained after repeating the process until reaching the last N levels, the maximum number of votes Either the answer or the image will be rejected. With the above processing, it is possible to expect significant efficiency for images whose processing is terminated at an early stage, and recognition over time can be performed as necessary.
- ⁇ 1> it is desirable to judge as accurately as possible with a small amount of calculation.
- ⁇ 2> is a measure for reducing the processing efficiency even for images that repeat the recognition process until the latter stage.
- the amount of computation when processing progresses to s stages with multistage classifiers should be the same as when a classifier with the same degree of approximation as stage s is used alone. . Each will be described below.
- t is the threshold for the number of votes
- r is the threshold for the ratio of the first and second votes.
- N approximate nearest neighbor searchers 1,..., ⁇ (hereinafter simply referred to as searchers) with different degrees of approximation. It is assumed that the degree of approximation is stronger for searcher (s-1) than for searcher s.
- P ( S ) is a set of feature vectors obtained as a result of distance calculation as a result of an approximate nearest neighbor search performed on feature vector q using searcher s. Approximate nearest neighbor search usually has the property that the stronger the degree of approximation, the smaller the number of feature vectors for distance calculation! That is,
- the searcher with monotonicity is used to construct the multi-stage classifier shown in Fig. 18, the difference P (S) — ⁇ ⁇ — ⁇ from the previous stage is not calculated in the s-th stage. It can be considered as a target of With this configuration, the union of feature vectors subject to distance calculation or voting from the first stage to the sth stage is the set P (S) when the searcher s is used alone. Therefore, the number of distance calculations or votes is the same. Furthermore, if the searcher has a differential search capability, the increase in the amount of calculation can be kept low even if multistage processing is performed.
- the block in the frame with the reference numeral 10 shows the detailed configuration of each stage classifier constituting the multistage classifier.
- the provisional nearest neighbor P * for each feature vector q is found and recorded in the provisional nearest neighbor database 16. Therefore, in the s-th stage, the differential feature vector pe (P ( s) — P ( s — D) is obtained by the differential hash key calculation, and the distance calculation with respect to q is performed only for them, and D If a near-field vector is found, it is newly registered in the provisional nearest neighbor database 16 as a provisional nearest neighbor 13 *, and voting is performed again.
- the blocks in the frame denoted by reference numeral 10 in FIG. 24 indicate the detailed configuration of each stage of the classifier that constitutes the multistage classifier.
- 23 and 24 includes the function of the image selection unit 23 shown in FIGS.
- the reliability judgment unit 24 determines whether the voting results up to the sth stage are sufficient. When sufficient reliability is obtained, the image to be recognized is determined (corresponding to the function of the image selection unit 23). However, if sufficient reliability is not obtained, it is determined that the next stage (s + 1) should be further advanced. If sufficient reliability is not obtained even when the final stage (Nth stage) is reached, it is determined that the result will be rejected.
- PCA-SIFT is used as a local descriptor.
- the inventors propose a method that is faster than conventional ANN and LSH as one aspect of the present invention.
- the nearest neighbor search method of the inventors has monotonicity and differential searchability as described later, and is very suitable for multi-stage. Therefore, in the multistage embodiment, a configuration in which the method of the inventors is applied as a discriminator will be described.
- the above-described multi-stage method is not necessarily limited to the combination with the inventors' method, and the conventional nearest neighbor search is acceptable as long as it satisfies the monotonicity and the difference searchability. It is thought that a certain effect can be obtained even with a classifier to which the method is applied.
- monotonicity can be satisfied by ANN and LSH.
- ANN monotonicity is satisfied when the value of allowable error ⁇ described later is changed stepwise
- LSH monotonicity is satisfied when the number L of hash tables to be searched is increased stepwise. That is, even with a conventional classifier, if it is multistaged, the object recognition processing time can be shortened compared to the case of one stage.
- the inventors' method to be applied to the discriminator need not necessarily be multi-staged. Even a one-stage classifier can reduce the processing time compared to a conventional classifier. However, if the classifier to which the inventors' method is applied is multi-staged, faster object recognition can be realized. Therefore, it is highly desirable to combine the two.
- the inventors' method applied to the classifier includes a method for performing distance calculation at the final stage of approximate nearest neighbor search (a method with distance calculation) and a method for eliminating distance calculation at all ( There is a method without distance calculation).
- a method with distance calculation a method for performing distance calculation at the final stage of approximate nearest neighbor search
- a method for eliminating distance calculation at all There is a method without distance calculation.
- the inventors disclose two speed-up methods using a hash table.
- One of the speed-up methods is a method of reducing the number of feature vector distance calculations. Specifically, when there are many feature vectors in the vicinity and many distance calculations are unavoidable, the speed is increased by discarding such feature vectors.
- this method is referred to as “with distance calculation”.
- the other is a method that does not perform any distance calculation. The only process is to draw a hash table and vote. Hereinafter, this method is referred to as “no distance calculation”.
- processing for recognizing an object from an image captured by a camera can shorten more than technology.
- processing can be performed with a smaller memory capacity than the conventional technology.
- the calculation time required to achieve the same recognition rate may be shorter than in the case of using the conventional approximate nearest neighbor search method such as ANN or LSH.
- the calculation time was reduced from 1/2 to 1/3 of the conventional technology. Also, the method without distance calculation is superior in terms of scalability because it uses less memory.
- Fig. 15 and 16 are block diagrams showing an example of the configuration of the image recognition apparatus of the present invention.
- Fig. 15 is a block diagram corresponding to the method without distance calculation
- Fig. 16 is a block diagram corresponding to the method with distance calculation.
- the image recognition method of the present invention is executed, for example, on the image recognition apparatus.
- the hardware of the image recognition device includes, for example, a CPU and a storage device such as a hard disk device that stores a program indicating the processing procedure executed by the CPU, a RAM that provides a work area for the CPU, and an input / output circuit that inputs and outputs data. Etc. More specifically, for example, a personal computer having the above configuration may be used. Alternatively, as a different mode, the device built-in type device may be composed of a large-scale integrated circuit (LSI), a hard disk device, and a microcomputer that controls processing thereof.
- LSI large-scale integrated circuit
- the feature point extraction unit 11 performs feature analysis from the pattern of the object included in the input image. This is a block for extracting a tuttle.
- the index calculation unit 13 is a block that calculates the index of the hash table from the feature vector by a predetermined calculation method.
- the image database 15 a plurality of images with image IDs are registered. Further, the image database 15 has a hash table 17 for referring to images.
- the image IDs of the images associated with the indexes are registered! /.
- the image ID is associated with each index as follows. First, a feature vector is extracted from the image to be registered by the same processing as the feature point extraction unit 11. For each extracted feature vector, the index of the hash table is calculated by the same calculation method as the index calculation unit 13. The image ID of the image including the feature vector for which the index is calculated is registered in advance for the calculated index.
- the voting unit 19 refers to the specific index of the hash table 17, and if there is an image ID registered in the hash table 17 for the referenced index, the voting unit 19 votes for that image It is. For voting, there is a voting table 21 that stores the number of votes for each image.
- the image selection unit 23 is a block that refers to the vote table 21 and selects an image that has obtained the maximum number of votes.
- the index calculation unit 13, the voting unit 19, and the voting table 21 are the target of multi-stage among the above-described blocks.
- the feature point extraction unit 11, the index calculation unit 13, the voting unit 19, the voting table 21, and the image selection unit 23 have the same functions as in FIG.
- the image database 35 is different from FIG. 15 in the configuration of the hash table 37. That is, in the index of the hash table 37, a vector element and an image ID of an image including the feature vector are registered in pairs for each feature vector of the registered image. The vector elements are used for distance calculations. Further, the image recognition apparatus in FIG. 16 includes a feature point matching unit 38.
- the feature point matching unit 38 when a plurality of feature vectors are registered for one index, This block calculates the distance between the vector and the feature vector extracted from the input image, determines the shortest distance, and determines the image ID registered with the shortest distance feature vector as a candidate image.
- the index calculation unit 13, the voting unit 19, the voting table 21, and the feature point matching unit 38 among the above-described blocks are multi-staged. It becomes a target.
- SIFT Scale-Invariant Feature Transform
- Lowe for example, DG Lowe, "Distinctive imag e features from scale-invariant keypoints, International Journal of Computer Vision, vol.60, no.2, pp.91-110, 2004.
- GPU Graphic Processing Unit
- feature points are extracted using software provided by Lowe (URL: http: ⁇ www.cs.web / Oowe / keypoints /).
- the feature vector is a 128-dimensional vector of integer values (0-255).
- PCA-SIFT that improves the stability and discriminability of SIF T by applying principal component analysis (PCA) to SIFT feature vectors (for example, Y. Ke and R. Sukthankar, Pca_sift: A more distinctive representation ror local image descriptors, and VPR2004, Vol.2, pp.506-513, 2004.).
- PCA-SIFT is used as an image local descriptor.
- the feature vector obtained by PCA-SIFT is a 36-dimensional real-valued vector. That is, for the feature vector obtained from SIFT power, URL: http: ⁇ ww Using the software provided in w.cs.cmu.edu/ke/pcasift/, it is converted into a 36-dimensional vector.
- Figure 1 is a graph showing the feature vector value distribution. The horizontal axis is the value of each dimension, and the vertical axis is the frequency.
- the first dimension is a bimodal distribution, and the second and subsequent dimensions are unimodal. Also, the variance decreases as the dimension increases. The average values are all near zero.
- each image represents one object.
- the object recognition task is defined as searching the database for an image that best matches the search question.
- a voting method is used in the present embodiment.
- the image of the search question is now Q, and the image in the database is p.
- the d-dimensional feature vectors obtained from Q and p are expressed as q and. If p is obtained as the feature vector corresponding to q as a result of the approximate nearest neighbor search, one vote is cast on image P.
- Such voting is performed for all feature vectors obtained from Q, and the image with the maximum number of votes is presented as the recognition result.
- ANN Approximate Nearest Neighbor listed in Non-Patent Document 3 is a technique for performing a nearest neighbor search at high speed using a tree structure.
- Tree nodes correspond to hyperrec tangles (hereinafter called cells) that divide the feature space, and feature vectors are also associated with leaf nodes.
- the search time can be reduced because the number of messages is reduced.
- LSH Location Sensitive Hashing listed in Non-Patent Document 4 is a method of nearest neighbor search using a hash table.
- E 2 LSH Exact Euclidean LSH; hereinafter simply referred to as LSH
- V is a function that converts V to an integer and has the following form.
- a is a d-dimensional vector in which each dimension is independently generated according to normal random numbers
- t Is a scalar defined by a uniform random number of [0, w].
- the present invention achieves high-speed processing compared to the case of using ANN or LSH by applying a large approximation.
- the speed-up problem is solved by using a hash function that has vitality in data characteristics.
- the following two methods are considered.
- One is a method of calculating the distance but reducing the number of target feature vectors. Specifically, when many collisions occur, that is, when many feature vectors having the same hash value are registered, they are deleted from the hash table in advance. This makes it possible to reduce the number of distance calculations per feature vector of the search query to a certain value or less.
- the other is a method that does not calculate distance at all. If erasure is performed according to the number of collisions, a feature vector effective for identifying an image remains in the hash table. Therefore, these features If you use this method, you can expect to get correct results by voting alone.
- the method of registering feature vectors in the hash table is as follows. Take the first through d-th dimensions of the 36-dimensional real-valued vector p obtained by PCA-SIFT,
- the hash is obtained by and is registered in the hash table.
- U is a discrete value Type (expressed in U base)
- H is the size of the hash table. Data to be registered in the hash table
- the data varies depending on whether distance is used.
- distance in addition to the image ID for the feature vector, p itself is registered and used for the distance calculation during the search. On the other hand, if distance is not used, registration of p is unnecessary.
- Table size The data registered in the hash table varies depending on whether distance is used. When using the distance, in addition to the image ID for the feature vector p, register p itself and use it for the distance calculation during the search. On the other hand, if distance is not used, registration of p is unnecessary.
- a threshold value c for the list length n is provided, and when n> c is satisfied, the entire list is deleted from the hash table.
- various weights used in information retrieval were also tested, and the recognition rate was not significantly different. Since the deletion is advantageous not only for the recognition rate but also for the speed, this embodiment employs deletion instead of weighting.
- the feature vector does not contribute much to image identification. Therefore, even if it is deleted, the impact is considered to be relatively small! /.
- Data registration is completed by performing the above processing on all feature vectors registered in the database.
- a feature vector is searched from the above hash table. If the set of obtained feature vectors is P, then the feature vector p * that is closest to q is found from P.
- the most important step in this process is the ability to search for feature vectors for les and cr.
- the simplest method is to obtain a bit vector for q as in the case of registration, and to obtain a feature vector having the same hash value using a hash function.
- the number of distance calculations can be reduced sufficiently, but a sufficient recognition rate cannot be obtained for the following reasons!
- the value of each dimension of the feature vector is a force S that changes depending on the shooting conditions. If there is a fluctuation that exceeds the threshold, the bit vector will be different and the corresponding feature vector can no longer be obtained.
- the value near the threshold is moved randomly by adding a uniform random number t to the value in equation (1).
- the relative position of the threshold is changed by applying a rotation matrix to the feature vector.
- the dimension number b to be processed is set to a value that is not so large.
- the discrete value used for index calculation may be selected at random, and only 1 and 2 may be used.
- the feature vector is searched. However, if such a process of “try both” is introduced without restriction, enormous calculation time is required. In this process, if the number of dimensions to be processed is b, the hash table is accessed using 2 b bit vectors. Therefore, in this embodiment, b is set to a value that is not so large.
- the method uses a distance meter as described above for the feature vector q of the search query. All feature vectors that belong to the set P of feature vectors obtained from the schew table.
- the parameter to be multistaged is not limited to b. Multiple stages with other parameters are possible.
- the parameter d satisfies not only monotonicity but also differential searchability.
- c and e There is the possibility of c and e.
- the method without distance calculation belongs to the difference set P (s) — P (s — 1 > obtained by voting while updating the provisional nearest neighbor p * in each stage of processing. Vote for all features.
- FIG. 4 is an explanatory diagram showing an example of a registered image used in the experiment.
- A is 3,100 images collected using Google image search. Search keywords such as posters, magazines, and covers were used.
- An example is shown in Fig. 4 (a).
- B is an image published on the PCA-SIFT site (URL: http: ⁇ www.cs.cmu.edu/ ⁇ yke/pcasift/), and the number of images is 18,500.
- This data is mainly composed of natural photographs and photographs of people.
- An example is shown in Fig. 4 (b).
- C consists of 78,400 images collected by tags such as a nimal, birthday, food, and japan on the photo sharing site flickr. Mainly includes objects, natural photographs, and human photographs as shown in Fig. 4 (c). During collection, images with a size of 600 X 600 pixels or less were excluded, and the image was reduced so that the long side of the image was 640 pixels or less. We also excluded images with 100 or fewer feature vectors. The average length of one side of the image was 498, 612, and 554 pixels, respectively.
- a large database includes a small database as part of it.
- An average of 2,069 feature vectors per image was extracted from DB3.
- the captured image was reduced to 512 X 341 pixels, and the feature vector was obtained by PCA-SIFT. As a result, an average of 605 feature vectors was obtained per image.
- OKI (Registered trademark) C5200n (color laser printer), CANON (registered trademark) EOS Kiss (registered trademark) Digital (6.3 million pixels) and attached lens EF-S 18_55mm USM were used for photographing.
- ANN and LSH were used as comparison methods for approximate nearest neighbor search and compared with the method of the present invention.
- ANN URL http: ⁇ www.cs.umd.edu/ ⁇ mount/ANN LSH URL: http: ⁇ ⁇ ⁇ ⁇ www.mit.edu/ ⁇ andoni/ Was used.
- the recognition rate and processing time were used as evaluation criteria.
- the recognition rate represents the rate at which search query images were correctly recognized.
- the processing time represents the time required for the search per image box of the search question. However, the time required for feature vector extraction is not included.
- the computer used in the experiment has a CPU power of S AMD Opteron (registered trademark) of 2.8 GHz and a memory of 16 GB.
- the threshold d for the maximum distance in the method with distance calculation is fixed at 3,000.
- Figure 6 shows the experimental results for the recognition rate and processing time when the tolerance ⁇ is varied from 2 to 100 using ANN. with increasing epsilon, recognition rate, that force s Such processing time is reduced.
- Figure 7 shows the experimental results for the recognition rate and processing time when the number of vector dimensions k and the number of hash functions L are changed using LSH.
- Figure 8 shows the experimental results of the recognition rate and processing time when c is changed to 24, 26, and 28. It can be seen that the processing time decreases as c decreases. However, the recognition rate decreased if c was too small. This is thought to be because the things that contributed to recognition were also deleted. On the other hand, when c was increased, the calculation time increased, but the recognition rate decreased little. This is because even if a feature vector that cannot be the nearest neighbor is searched, it can be excluded by distance calculation.
- the horizontal axis represents the recognition rate
- the vertical axis Figure 12 shows a graph depicting the processing time.
- An ANN with parameters changed was drawn as a line and used as the evaluation standard. The higher the recognition rate is, the shorter the processing time is. Therefore, it can be said that the one plotted in the lower right is superior. LSH never crossed the ANN line.
- the method of calculating the distance by the method of the present invention was superior to ANN when the recognition rate was 98% or less.
- the method of the present invention that does not calculate distance was superior to ANN in most cases.
- Figure 13 shows the processing time of about 10 ms and the best recognition rate.
- Table 2 shows the recognition rate and processing time using typical values of various parameters.
- the method with distance calculation achieves the same recognition rate as the ANN in about 1/3 of the processing time.
- the average recognition rate does not reach ANN.
- the recognition rate of about 96% is 4ms or less! /, Short! /, And it can be realized in processing time.
- the method that does not use the distance calculation is an excellent method in terms of scalability when it can be satisfied with a certain degree of recognition rate, although it is inferior in terms of the recognition rate.
- the processing is as simple as accessing the hash table and voting, there may be advantages in this aspect.
- ANN and LSH which are conventional methods of approximate nearest neighbor search
- a one-stage classifier is constructed using a method with distance calculation, and a multi-stage classifier using the inventors' technique is used. Compared with the case.
- the local descriptor provided by the PCA-SIFT site was used.
- the processing time shown below is for recognition per image of the search question.
- the images in the image database were collected from the same source as in Experiment 1.
- the number of images collected was 3,100 images collected using Google's image search.
- Image power published on the PCA-SIFT site 3 ⁇ 4 450 images collected by tags such as animal, birthday, food on the photo sharing site flickr, 3,450 images, a total of 10,000 images.
- Two types of search questions were created: those with and without images corresponding to the database.
- a total of 500 images of 100, 200, and 200 images were randomly selected from the images included in the database for each collection method.
- they were printed on A4 paper and photographed using a camera.
- the camera's optical axis angle ⁇ relative to the paper surface is 90 °, 75 in an arrangement that captures the entire paper surface. , 60 °.
- a part of the paper was photographed at an angle of 90 °.
- a total of four images were obtained on the paper surface of the kite.
- the captured image was reduced to 512 X 341 pixels, and feature vectors were obtained by PCA-SIFT. As a result, an average of 612 feature vectors per image ⁇ was obtained.
- the accuracy of the nearest neighbor search necessary for image recognition was examined. Specifically, we measured the relationship between the accuracy of approximate nearest neighbor search and the recognition rate of the image by changing the parameters for the method without multi-step (method with ANN, LSH, distance calculation). The accuracy of the approximate nearest neighbor search is the rate at which the true nearest neighbor is found by the approximate nearest neighbor search.
- a graph depicting the relationship between recognition rate and processing time with various parameters is shown in Fig. 22.
- An ANN with an allowable error ⁇ changed is drawn as a line and used as a standard for evaluation. The higher the recognition rate is, the shorter the processing time is. Therefore, it can be said that the one plotted in the lower right is superior. LSH never crossed the ANN line.
- the maximum recognition rate does not reach that of ANN. It has been realized.
- the processing time is reduced to about 1/5 of the method with distance calculation by multi-stepping.
- the evaluation scale is defined as follows. For search queries with corresponding images, recognition rate C, false recognition rate E, rejection rate R
- the object recognition process using the present invention can be applied to a service that recognizes an object from an image captured by a camera and performs information processing according to the recognition result.
- information processing there is a process of indexing an existing image or video image.
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- General Physics & Mathematics (AREA)
- Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Computation (AREA)
- Multimedia (AREA)
- Databases & Information Systems (AREA)
- Software Systems (AREA)
- Medical Informatics (AREA)
- General Health & Medical Sciences (AREA)
- Computing Systems (AREA)
- Health & Medical Sciences (AREA)
- Data Mining & Analysis (AREA)
- Life Sciences & Earth Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- General Engineering & Computer Science (AREA)
- Image Analysis (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
- Processing Or Creating Images (AREA)
Description
Claims
Priority Applications (5)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
EP07791766A EP2063394A4 (en) | 2006-08-31 | 2007-08-01 | IMAGE RECOGNITION METHOD, IMAGE RECOGNITION DEVICE, AND IMAGE RECOGNITION PROGRAM |
US12/439,493 US8199973B2 (en) | 2006-08-31 | 2007-08-01 | Image recognition method, image recognition device, and image recognition program |
JP2008532003A JP4883649B2 (ja) | 2006-08-31 | 2007-08-01 | 画像認識方法、画像認識装置および画像認識プログラム |
CN200780040849XA CN101536035B (zh) | 2006-08-31 | 2007-08-01 | 图像识别方法、以及图像识别装置 |
HK10102372.2A HK1134364A1 (en) | 2006-08-31 | 2010-03-05 | Image recognition method and image recognition device |
Applications Claiming Priority (4)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
JP2006236113 | 2006-08-31 | ||
JP2006-236113 | 2006-08-31 | ||
JP2007129737 | 2007-05-15 | ||
JP2007-129737 | 2007-05-15 |
Publications (1)
Publication Number | Publication Date |
---|---|
WO2008026414A1 true WO2008026414A1 (fr) | 2008-03-06 |
Family
ID=39135700
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/JP2007/065086 WO2008026414A1 (fr) | 2006-08-31 | 2007-08-01 | Procédé de reconnaissance d'image, dispositif de reconnaissance d'image et programme de reconnaissance d'image |
Country Status (6)
Country | Link |
---|---|
US (1) | US8199973B2 (ja) |
EP (1) | EP2063394A4 (ja) |
JP (1) | JP4883649B2 (ja) |
CN (1) | CN101536035B (ja) |
HK (1) | HK1134364A1 (ja) |
WO (1) | WO2008026414A1 (ja) |
Cited By (16)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2009133856A1 (ja) | 2008-04-28 | 2009-11-05 | 公立大学法人大阪府立大学 | 物体認識用画像データベースの作成方法、処理装置および処理用プログラム |
WO2010101186A1 (ja) | 2009-03-04 | 2010-09-10 | 公立大学法人大阪府立大学 | 画像検索方法、画像検索プログラム及び画像登録方法 |
WO2010143573A1 (ja) | 2009-06-10 | 2010-12-16 | 公立大学法人大阪府立大学 | 物体認識用画像データベースの作成方法、作成装置および作成処理プログラム |
WO2011092865A1 (ja) * | 2010-02-01 | 2011-08-04 | 株式会社モルフォ | 物体検出装置及び物体検出方法 |
WO2011136276A1 (ja) * | 2010-04-28 | 2011-11-03 | 公立大学法人大阪府立大学 | 三次元物体認識用画像データベースの作成方法および作成装置 |
JP2012043437A (ja) * | 2010-08-17 | 2012-03-01 | Fujitsu Ltd | 画像処理方法及び画像処理装置 |
JP2012160047A (ja) * | 2011-02-01 | 2012-08-23 | Denso It Laboratory Inc | 対応参照画像検索装置及び方法、コンテンツ重畳装置、システム、及び方法、並びにコンピュータプログラム |
CN102722554A (zh) * | 2012-05-28 | 2012-10-10 | 中国人民解放军信息工程大学 | 位置敏感哈希随机性减弱方法 |
US8306315B2 (en) | 2008-04-30 | 2012-11-06 | Osaka Prefecture University Public Corporation | Method of compiling three-dimensional object identifying image database, processing apparatus and processing program |
JP2012531130A (ja) * | 2009-06-26 | 2012-12-06 | インテル・コーポレーション | ビデオコピーを検知する技術 |
CN103678660A (zh) * | 2013-12-24 | 2014-03-26 | 北京邮电大学 | 一种图像检索方法 |
JP2014203289A (ja) * | 2013-04-05 | 2014-10-27 | キヤノン株式会社 | ハッシュ値生成装置、システム、判定方法、プログラム、記憶媒体 |
US9116961B2 (en) | 2011-05-06 | 2015-08-25 | Fujitsu Limited | Information processing device, information processing system and search method |
JP2017049815A (ja) * | 2015-09-02 | 2017-03-09 | トヨタ自動車株式会社 | 物体認識方法、物体認識装置及びプログラム |
JPWO2017072890A1 (ja) * | 2015-10-28 | 2018-05-17 | 株式会社東芝 | データ管理システム、データ管理方法およびプログラム |
WO2021025092A1 (ja) * | 2019-08-06 | 2021-02-11 | 学校法人明星学苑 | 圧縮データ検索エンジンを用いた情報処理装置およびその情報処理方法 |
Families Citing this family (49)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JP2009129337A (ja) * | 2007-11-27 | 2009-06-11 | Hitachi Ltd | 三次元類似形状検索装置 |
US8140448B2 (en) | 2008-05-09 | 2012-03-20 | International Business Machines Corporation | System and method for classifying data streams with very large cardinality |
US8429216B2 (en) * | 2008-09-23 | 2013-04-23 | Hewlett-Packard Development Company, L.P. | Generating a hash value from a vector representing a data object |
JP5714599B2 (ja) * | 2009-12-02 | 2015-05-07 | クゥアルコム・インコーポレイテッドQualcomm Incorporated | イメージ認識のための記述子パッチの高速部分空間射影 |
US8352494B1 (en) * | 2009-12-07 | 2013-01-08 | Google Inc. | Distributed image search |
CN101790064A (zh) * | 2010-02-23 | 2010-07-28 | 公安部第三研究所 | 具有视频结构化描述功能的硬盘录像设备及其方法 |
US9058561B2 (en) | 2010-03-22 | 2015-06-16 | Aptima, Inc. | Systems and methods of cognitive patterns knowledge generation |
US9530073B2 (en) | 2010-04-20 | 2016-12-27 | Qualcomm Incorporated | Efficient descriptor extraction over multiple levels of an image scale space |
JP5782037B2 (ja) * | 2010-09-10 | 2015-09-24 | パナソニック インテレクチュアル プロパティ コーポレーション オブアメリカPanasonic Intellectual Property Corporation of America | 排他的分類器による一般物体の画像認識装置及び方法 |
US8548237B2 (en) | 2010-10-18 | 2013-10-01 | Hewlett-Packard Development Company, L.P. | Ordinal and spatial local feature vector based image representation |
KR101675785B1 (ko) | 2010-11-15 | 2016-11-14 | 삼성전자주식회사 | 특징점을 이용한 영상 검색 방법 및 상기 방법을 수행하는 장치 |
US8612441B2 (en) * | 2011-02-04 | 2013-12-17 | Kodak Alaris Inc. | Identifying particular images from a collection |
US9122705B1 (en) * | 2011-03-15 | 2015-09-01 | Google Inc. | Scoring hash functions |
JP5692725B2 (ja) * | 2011-05-27 | 2015-04-01 | 公立大学法人大阪府立大学 | 近似最近傍探索に係るデータベースの登録方法および登録装置 |
CN102855498B (zh) * | 2011-07-01 | 2016-08-31 | 富士通株式会社 | 字符识别方法和装置 |
JP5485954B2 (ja) * | 2011-09-06 | 2014-05-07 | 東芝テック株式会社 | 店舗システム及びプログラム |
US9400945B2 (en) * | 2011-09-23 | 2016-07-26 | GM Global Technology Operations LLC | System and method of fast object detection using parts to whole fragment detection |
US10152674B2 (en) | 2012-01-16 | 2018-12-11 | Texas Instruments Incorporated | Accelerated decision tree execution |
CN103377373A (zh) * | 2012-04-25 | 2013-10-30 | 佳能株式会社 | 图像特征产生方法及设备、分类器、系统和捕获设备 |
US9165068B2 (en) * | 2012-08-03 | 2015-10-20 | Adobe Systems Incorporated | Techniques for cloud-based similarity searches |
CN102930529B (zh) * | 2012-09-24 | 2014-04-16 | 南京工业大学 | 一种基于非对称模板搜索的快速伪造区域盲检测方法 |
IL226219A (en) * | 2013-05-07 | 2016-10-31 | Picscout (Israel) Ltd | Efficient comparison of images for large groups of images |
US9286549B1 (en) | 2013-07-15 | 2016-03-15 | Google Inc. | Sublinear time classification via feature padding and hashing |
JP5808371B2 (ja) * | 2013-08-28 | 2015-11-10 | ヤフー株式会社 | 画像認識装置、画像認識方法及び画像認識プログラム |
JP6151141B2 (ja) * | 2013-09-18 | 2017-06-21 | 株式会社東芝 | 仕分装置および仕分方法 |
US9569692B2 (en) * | 2014-10-31 | 2017-02-14 | The Nielsen Company (Us), Llc | Context-based image recognition for consumer market research |
US9600524B2 (en) | 2014-12-22 | 2017-03-21 | Blackberry Limited | Method and system for efficient feature matching |
CN104732221A (zh) * | 2015-03-30 | 2015-06-24 | 郑州师范学院 | 一种基于OpenCL并行加速的SIFT特征匹配方法 |
CN104933407A (zh) * | 2015-05-28 | 2015-09-23 | 成都佳发安泰科技股份有限公司 | 基于sift变换的指纹识别方法 |
US10885098B2 (en) | 2015-09-15 | 2021-01-05 | Canon Kabushiki Kaisha | Method, system and apparatus for generating hash codes |
US10424072B2 (en) | 2016-03-01 | 2019-09-24 | Samsung Electronics Co., Ltd. | Leveraging multi cues for fine-grained object classification |
US10579860B2 (en) | 2016-06-06 | 2020-03-03 | Samsung Electronics Co., Ltd. | Learning model for salient facial region detection |
US10740646B2 (en) * | 2017-10-30 | 2020-08-11 | The Aerospace Corporation | Programmable cellular automata for memory search, recall, simulation, and improvisation |
US10832180B2 (en) | 2017-10-30 | 2020-11-10 | The Aerospace Corporation | Artificial intelligence system that employs windowed cellular automata to create plausible alternatives |
US10762431B2 (en) | 2017-10-30 | 2020-09-01 | The Aerospace Corporation | Low probability transitions and boundary crossing into disallowed states for a more optimal solution |
CN108805157B (zh) * | 2018-04-11 | 2022-02-18 | 南京理工大学 | 基于部分随机监督离散式哈希的遥感图像分类方法 |
JP7207862B2 (ja) * | 2018-04-26 | 2023-01-18 | 株式会社日立製作所 | 物体認識装置および方法 |
CN110471942B (zh) * | 2018-05-08 | 2022-04-15 | 北京大学 | 一种基于均衡哈希编码的飞行器查询方法及系统 |
US11120070B2 (en) * | 2018-05-21 | 2021-09-14 | Microsoft Technology Licensing, Llc | System and method for attribute-based visual search over a computer communication network |
CN109389148B (zh) * | 2018-08-28 | 2021-11-23 | 昆明理工大学 | 一种基于改进DHash算法的图像相似判定方法 |
EP3899789A4 (en) * | 2018-12-20 | 2022-10-05 | Tracxone Ltd. | SYSTEM AND METHOD FOR CLASSIFIER LEARNING AND RECOVERY FROM A CLASSIFIER DATABASE FOR LARGE-SCALE PRODUCT IDENTIFICATION |
CN110134804B (zh) * | 2019-05-20 | 2021-09-17 | 北京达佳互联信息技术有限公司 | 图像检索方法、装置及存储介质 |
CN112529921A (zh) * | 2019-09-24 | 2021-03-19 | 上海壹品仓软件科技有限公司 | 一种线上线下多渠道商品库存共享系统和方法 |
CN110751188B (zh) * | 2019-09-26 | 2020-10-09 | 华南师范大学 | 基于多标记学习的用户标签预测方法、系统及存储介质 |
CN111008210B (zh) * | 2019-11-18 | 2023-08-11 | 浙江大华技术股份有限公司 | 商品识别方法、装置、编解码器及存储装置 |
US12087027B2 (en) * | 2020-01-31 | 2024-09-10 | Nec Corporation | Object recognition apparatus, object recognition method, and recording medium |
US11645733B2 (en) | 2020-06-16 | 2023-05-09 | Bank Of America Corporation | System and method for providing artificial intelligence architectures to people with disabilities |
WO2022093263A1 (en) * | 2020-10-30 | 2022-05-05 | Hewlett-Packard Development Company, L.P. | Multi-modal hierarchical semantic search engine |
US20240078270A1 (en) * | 2022-09-07 | 2024-03-07 | Sage Global Services Limited | Classifying documents using geometric information |
Citations (10)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JPS60142788A (ja) * | 1983-12-29 | 1985-07-27 | Hitachi Ltd | パターン認識における特微量評価方法および装置 |
JPS63132381A (ja) * | 1986-11-25 | 1988-06-04 | Hitachi Ltd | 画像デ−タ検索・表示システム |
JPH0484277A (ja) * | 1990-07-26 | 1992-03-17 | Nec Corp | 特徴量選択方法及び装置と高速識別方法及び装置 |
JPH05233881A (ja) * | 1992-02-21 | 1993-09-10 | Fujitsu Ltd | 文字認識装置における認識文字テーブル作成方法 |
JPH06168277A (ja) * | 1992-11-30 | 1994-06-14 | Canon Inc | 画像検索装置及び方法 |
JP2000285190A (ja) * | 1999-03-31 | 2000-10-13 | Toshiba Corp | 帳票識別方法および帳票識別装置および記憶媒体 |
JP2003242509A (ja) * | 2001-12-13 | 2003-08-29 | Toshiba Corp | パターン認識装置及びその方法 |
JP2004133629A (ja) * | 2002-10-09 | 2004-04-30 | Ricoh Co Ltd | 特定マーク検出用辞書作成装置、特定マーク検出装置、特定マーク認識装置並びにプログラムおよび記録媒体 |
US20040208372A1 (en) | 2001-11-05 | 2004-10-21 | Boncyk Wayne C. | Image capture and identification system and process |
WO2006092957A1 (ja) | 2005-03-01 | 2006-09-08 | Osaka Prefecture University Public Corporation | 文書・画像検索方法とそのプログラム、文書・画像登録装置および検索装置 |
Family Cites Families (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US4658429A (en) * | 1983-12-29 | 1987-04-14 | Hitachi, Ltd. | System and method for preparing a recognition dictionary |
US7200270B2 (en) * | 2001-12-13 | 2007-04-03 | Kabushiki Kaisha Toshiba | Pattern recognition apparatus and method using distributed model representation of partial images |
US7684029B2 (en) * | 2004-10-29 | 2010-03-23 | Avago Technologies General Ip (Singapore) Pte. Ltd. | Method and apparatus for identifying a sensed light environment |
-
2007
- 2007-08-01 JP JP2008532003A patent/JP4883649B2/ja not_active Expired - Fee Related
- 2007-08-01 WO PCT/JP2007/065086 patent/WO2008026414A1/ja active Application Filing
- 2007-08-01 CN CN200780040849XA patent/CN101536035B/zh not_active Expired - Fee Related
- 2007-08-01 US US12/439,493 patent/US8199973B2/en not_active Expired - Fee Related
- 2007-08-01 EP EP07791766A patent/EP2063394A4/en not_active Withdrawn
-
2010
- 2010-03-05 HK HK10102372.2A patent/HK1134364A1/xx not_active IP Right Cessation
Patent Citations (10)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JPS60142788A (ja) * | 1983-12-29 | 1985-07-27 | Hitachi Ltd | パターン認識における特微量評価方法および装置 |
JPS63132381A (ja) * | 1986-11-25 | 1988-06-04 | Hitachi Ltd | 画像デ−タ検索・表示システム |
JPH0484277A (ja) * | 1990-07-26 | 1992-03-17 | Nec Corp | 特徴量選択方法及び装置と高速識別方法及び装置 |
JPH05233881A (ja) * | 1992-02-21 | 1993-09-10 | Fujitsu Ltd | 文字認識装置における認識文字テーブル作成方法 |
JPH06168277A (ja) * | 1992-11-30 | 1994-06-14 | Canon Inc | 画像検索装置及び方法 |
JP2000285190A (ja) * | 1999-03-31 | 2000-10-13 | Toshiba Corp | 帳票識別方法および帳票識別装置および記憶媒体 |
US20040208372A1 (en) | 2001-11-05 | 2004-10-21 | Boncyk Wayne C. | Image capture and identification system and process |
JP2003242509A (ja) * | 2001-12-13 | 2003-08-29 | Toshiba Corp | パターン認識装置及びその方法 |
JP2004133629A (ja) * | 2002-10-09 | 2004-04-30 | Ricoh Co Ltd | 特定マーク検出用辞書作成装置、特定マーク検出装置、特定マーク認識装置並びにプログラムおよび記録媒体 |
WO2006092957A1 (ja) | 2005-03-01 | 2006-09-08 | Osaka Prefecture University Public Corporation | 文書・画像検索方法とそのプログラム、文書・画像登録装置および検索装置 |
Non-Patent Citations (11)
Title |
---|
"Nearest-neighbor methods in learning and vision", 2005, THE MIT PRESS |
D.G. LOWE: "Distinctive image features from scale-invariant keypoints", INTERNATIONAL JOURNAL OF COMPUTER VISION, vol. 60, no. 2, 2004, pages 91 - 110, XP019216426, DOI: doi:10.1023/B:VISI.0000029664.99615.94 |
KATAYAMA NORIO; SATO SHINICHI: "Indexing Technique for Similarity Retrieval", JOURNAL OF INFORMATION PROCESSING SOCIETY OF JAPAN, vol. 42, no. 10, October 2001 (2001-10-01), pages 958 - 964 |
KOBAYASHI TAKUO; NAKAGAWA MASAKI: "Higher-dimensional Nearest Neighbor Search by Distributed Coding", IEICET TECHNICAL REPORT PRMU2006, June 2006 (2006-06-01), pages 41 |
M. DATAR ET AL.: "Locality-sensitive hashing scheme based on p-stable distributions", PROC. OF THE 20TH ANNUAL SYMPOSIUM ON COMPUTATIONAL GEOMETRY, 2004, pages 253 - 262, XP058169711, DOI: doi:10.1145/997817.997857 |
P. LNDYK: "Handbook of discrete and computational geometry", 2004, CHAPMAN & HALL/CRC, article "Nearest neighbors in high-dimensional spaces", pages: 877 - 892 |
S. ARYA ET AL.: "An optimal algorithm for approximate nearest neighbor searching", JOURNAL OF THE ACM, vol. 45, no. 6, 1998, pages 891 - 923, XP058146321, DOI: doi:10.1145/293347.293348 |
See also references of EP2063394A4 |
SHIBATA TOMOYUKI; KATO TAKEKAZU; WADA TOSHIKAZU: "K-D Decision tree: An Accelerated and Memory Efficient Nearest Neighbor Classifier", IEICE TRANSACTIONS (D-II), vol. J88-D-II, no. 8, August 2005 (2005-08-01), pages 1367 - 1377 |
WADA TOSHIKAZU: "Acceleration Method for Nearest Neighbor Classification based on Space Decomposition", JOURNAL OF INFORMATION PROCESSING SOCIETY OF JAPAN, vol. 46, no. 8, August 2005 (2005-08-01), pages 912 - 918 |
Y. KE; R. SUKTHANKAR: "Pca-sift: A more distinctive representation for local image descriptors", CVPR2004, vol. 2, 2004, pages 506 - 513 |
Cited By (31)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2009133856A1 (ja) | 2008-04-28 | 2009-11-05 | 公立大学法人大阪府立大学 | 物体認識用画像データベースの作成方法、処理装置および処理用プログラム |
JP5294342B2 (ja) * | 2008-04-28 | 2013-09-18 | 公立大学法人大阪府立大学 | 物体認識用画像データベースの作成方法、処理装置および処理用プログラム |
KR101247891B1 (ko) * | 2008-04-28 | 2013-03-26 | 고리츠다이가쿠호징 오사카후리츠다이가쿠 | 물체 인식용 화상 데이터베이스의 작성 방법, 처리 장치 및 처리용 프로그램 |
US8340451B2 (en) | 2008-04-28 | 2012-12-25 | Osaka Prefecture University Public Corporation | Method for constructing image database for object recognition, processing apparatus and processing program |
JP5278881B2 (ja) * | 2008-04-30 | 2013-09-04 | 公立大学法人大阪府立大学 | 3次元物体認識用画像データベースの作成方法、処理装置および処理用プログラム |
KR101257999B1 (ko) * | 2008-04-30 | 2013-04-24 | 고리츠다이가쿠호징 오사카후리츠다이가쿠 | 3차원 물체 인식용 화상 데이터베이스의 작성 방법, 처리 장치 및 처리용 프로그램 |
US8306315B2 (en) | 2008-04-30 | 2012-11-06 | Osaka Prefecture University Public Corporation | Method of compiling three-dimensional object identifying image database, processing apparatus and processing program |
CN102422319A (zh) * | 2009-03-04 | 2012-04-18 | 公立大学法人大阪府立大学 | 图像检索方法、图像检索程序和图像登记方法 |
US8818103B2 (en) | 2009-03-04 | 2014-08-26 | Osaka Prefecture University Public Corporation | Image retrieval method, image retrieval program, and image registration method |
CN102422319B (zh) * | 2009-03-04 | 2014-04-30 | 公立大学法人大阪府立大学 | 图像检索方法和图像存储方法 |
WO2010101186A1 (ja) | 2009-03-04 | 2010-09-10 | 公立大学法人大阪府立大学 | 画像検索方法、画像検索プログラム及び画像登録方法 |
WO2010143573A1 (ja) | 2009-06-10 | 2010-12-16 | 公立大学法人大阪府立大学 | 物体認識用画像データベースの作成方法、作成装置および作成処理プログラム |
JP2012531130A (ja) * | 2009-06-26 | 2012-12-06 | インテル・コーポレーション | ビデオコピーを検知する技術 |
US8693791B2 (en) | 2010-02-01 | 2014-04-08 | Morpho, Inc. | Object detection apparatus and object detection method |
CN102216958A (zh) * | 2010-02-01 | 2011-10-12 | 株式会社摩如富 | 物体检测装置以及物体检测方法 |
WO2011092865A1 (ja) * | 2010-02-01 | 2011-08-04 | 株式会社モルフォ | 物体検出装置及び物体検出方法 |
WO2011136276A1 (ja) * | 2010-04-28 | 2011-11-03 | 公立大学法人大阪府立大学 | 三次元物体認識用画像データベースの作成方法および作成装置 |
JP5818327B2 (ja) * | 2010-04-28 | 2015-11-18 | オリンパス株式会社 | 三次元物体認識用画像データベースの作成方法および作成装置 |
US8971610B2 (en) | 2010-04-28 | 2015-03-03 | Osaka Prefecture University Public Corporation | Method and apparatus of compiling image database for three-dimensional object recognition |
JP2012043437A (ja) * | 2010-08-17 | 2012-03-01 | Fujitsu Ltd | 画像処理方法及び画像処理装置 |
JP2012160047A (ja) * | 2011-02-01 | 2012-08-23 | Denso It Laboratory Inc | 対応参照画像検索装置及び方法、コンテンツ重畳装置、システム、及び方法、並びにコンピュータプログラム |
US9116961B2 (en) | 2011-05-06 | 2015-08-25 | Fujitsu Limited | Information processing device, information processing system and search method |
CN102722554B (zh) * | 2012-05-28 | 2014-07-02 | 中国人民解放军信息工程大学 | 位置敏感哈希随机性减弱方法 |
CN102722554A (zh) * | 2012-05-28 | 2012-10-10 | 中国人民解放军信息工程大学 | 位置敏感哈希随机性减弱方法 |
JP2014203289A (ja) * | 2013-04-05 | 2014-10-27 | キヤノン株式会社 | ハッシュ値生成装置、システム、判定方法、プログラム、記憶媒体 |
CN103678660A (zh) * | 2013-12-24 | 2014-03-26 | 北京邮电大学 | 一种图像检索方法 |
CN103678660B (zh) * | 2013-12-24 | 2017-01-11 | 北京邮电大学 | 一种图像检索方法 |
JP2017049815A (ja) * | 2015-09-02 | 2017-03-09 | トヨタ自動車株式会社 | 物体認識方法、物体認識装置及びプログラム |
JPWO2017072890A1 (ja) * | 2015-10-28 | 2018-05-17 | 株式会社東芝 | データ管理システム、データ管理方法およびプログラム |
US11281645B2 (en) | 2015-10-28 | 2022-03-22 | Kabushiki Kaisha Toshiba | Data management system, data management method, and computer program product |
WO2021025092A1 (ja) * | 2019-08-06 | 2021-02-11 | 学校法人明星学苑 | 圧縮データ検索エンジンを用いた情報処理装置およびその情報処理方法 |
Also Published As
Publication number | Publication date |
---|---|
HK1134364A1 (en) | 2010-04-23 |
JPWO2008026414A1 (ja) | 2010-01-14 |
JP4883649B2 (ja) | 2012-02-22 |
US20100027895A1 (en) | 2010-02-04 |
EP2063394A4 (en) | 2011-08-03 |
EP2063394A1 (en) | 2009-05-27 |
CN101536035B (zh) | 2012-09-26 |
US8199973B2 (en) | 2012-06-12 |
CN101536035A (zh) | 2009-09-16 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
WO2008026414A1 (fr) | Procédé de reconnaissance d'image, dispositif de reconnaissance d'image et programme de reconnaissance d'image | |
US11048966B2 (en) | Method and device for comparing similarities of high dimensional features of images | |
US10235623B2 (en) | Accurate tag relevance prediction for image search | |
Tolias et al. | Visual query expansion with or without geometry: refining local descriptors by feature aggregation | |
JP5294342B2 (ja) | 物体認識用画像データベースの作成方法、処理装置および処理用プログラム | |
US20210158164A1 (en) | Finding k extreme values in constant processing time | |
EP2284791B1 (en) | Method of creating three-dimensional object identifying image database, processing apparatus and processing program | |
WO2013129580A1 (ja) | 近似最近傍探索装置、近似最近傍探索方法およびそのプログラム | |
US20120084305A1 (en) | Compiling method, compiling apparatus, and compiling program of image database used for object recognition | |
Jégou et al. | Exploiting descriptor distances for precise image search | |
US12118813B2 (en) | Continuous learning for document processing and analysis | |
Yang et al. | Smart library: Identifying books on library shelves using supervised deep learning for scene text reading | |
Sadovnik et al. | Finding your lookalike: Measuring face similarity rather than face identity | |
Tian et al. | Image classification based on the combination of text features and visual features | |
Amato et al. | Aggregating binary local descriptors for image retrieval | |
JP6017277B2 (ja) | 特徴ベクトルの集合で表されるコンテンツ間の類似度を算出するプログラム、装置及び方法 | |
JP5833499B2 (ja) | 高次元の特徴ベクトル集合で表現されるコンテンツを高精度で検索する検索装置及びプログラム | |
Zhang et al. | Neural ranking for automatic image annotation | |
CN113920303B (zh) | 一种基于卷积神经网络弱监督类别无关的图像相似性检索系统及其控制方法 | |
Xu et al. | Learning multi-task local metrics for image annotation | |
Weng et al. | Random VLAD based deep hashing for efficient image retrieval | |
Voloshynovskiy et al. | On accuracy, robustness, and security of bag-of-word search systems | |
Lulu et al. | TMNIO: Triplet merged network with involution operators for improved few‐shot image classification | |
Maier et al. | Image auto-annotation with automatic selection of the annotation length | |
da Costa et al. | Deep convolutional features for fingerprint indexing |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
WWE | Wipo information: entry into national phase |
Ref document number: 200780040849.X Country of ref document: CN |
|
121 | Ep: the epo has been informed by wipo that ep was designated in this application |
Ref document number: 07791766 Country of ref document: EP Kind code of ref document: A1 |
|
WWE | Wipo information: entry into national phase |
Ref document number: 2008532003 Country of ref document: JP |
|
NENP | Non-entry into the national phase |
Ref country code: DE |
|
NENP | Non-entry into the national phase |
Ref country code: RU |
|
WWE | Wipo information: entry into national phase |
Ref document number: 2007791766 Country of ref document: EP |
|
WWE | Wipo information: entry into national phase |
Ref document number: 12439493 Country of ref document: US |