JP2015032220A - Image identification device and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide an image identification device capable of identifying an image accurately without receiving limitation due to color scheme of an identification object.SOLUTION: A mapping part 2 applies mapping by mapping factor to an input image for generating a mapping pixel signal. A calculation part 3 calculates a local feature amount of the mapping pixel signal. A collation part 4 collates the calculated local feature amount and a local feature amount of the identification object, for identifying whether the input image corresponds to the identification object or not. At the time, the applied mapping factor is determined as factor for reducing correlation of the signals by the image of the identification object. The local feature amount of the identification object is calculated in the mapping pixel signal in which the mapping based on the determined mapping factor is applied, to the image of the identification object.

Description

  The present invention relates to an image identification device and a program, and more particularly to an image identification device and a program for calculating a mapping coefficient for reducing a correlation between signals from an image to be identified and applying it to the image to be identified with high accuracy. Is.

  As a method for improving the identification performance in the conventional image identification, there is a method of calculating / collating the feature quantity of the identification target.

  As typical feature quantities, SIFT (Scale Invariant Feature Transform) and SURF (Speeded Up Robust Features) are known, and these have features that are robust to enlargement / reduction and rotation of the subject.

  Regarding the matching of the feature amounts, outlier removal can be performed to some extent by performing robust estimation such as RANSAC (Random Sample Consensus).

  In Patent Document 1, a color space that is resistant to illumination fluctuations is determined in advance as a color space to be converted by the color space conversion unit, thereby reducing the influence on the feature amount due to the difference in color, and a more appropriate feature amount. Disclosed are techniques that can be obtained.

JP 2012-146221 A

David G. Lowe, "Distinctive image features from scale-invariant key- points," Int. Journal of Computer Vision, Vol.60, No.2, pp.91-110, 2004. H. Bay, T. Tuytelaars, and L. Van Gool, "SURF: Speeded Up Robust Features," In ECCV, pp.404-417, 2006.

  In Non-Patent Document 1 and Non-Patent Document 2, since the color space is determined in advance, there is a problem in that feature points and feature amounts cannot be calculated depending on the color scheme of the identification target.

  Although Patent Document 1 considers color space conversion, it only mentions whether or not to use color space conversion that is robust to illumination fluctuations, and thus cannot affect the effect on the feature amount due to color differences other than illumination fluctuations. There is a problem. Further, the problem that the feature points and the feature amounts cannot be calculated depending on the color scheme of the identification target has not been solved.

  An object of the present invention is to solve the above-described problems of the prior art and to provide an image identification device and program having high identification performance.

  In order to achieve the above object, the present invention provides an image identification device for identifying whether an input image corresponds to an identification object, and applying a mapping based on a mapping coefficient to the input image to generate a mapped pixel signal The input image corresponds to the identification target by collating the calculated local feature quantity with the local feature quantity of the identification target. A matching unit for identifying whether the mapping coefficient to be applied is a mapping coefficient obtained by reducing the correlation between the signals of the image to be identified, and the local area of the identification target. The feature amount is calculated in a mapping pixel signal obtained by applying mapping based on the obtained mapping coefficient to the image to be identified.

  In addition, the present invention is an image identification program for causing a computer to function as the image identification device.

  According to the present invention, for a pixel signal to be identified configured in a predetermined color space, a mapping coefficient to be converted into a mapping pixel signal that reduces redundancy between signals is obtained in advance, and the mapping coefficient is calculated. By applying it in common to the identification target and the input image, the amount of information is concentrated on a part of the mapped pixel signal, which makes it easy to generate feature amounts and also causes differences between the feature amounts. Will improve.

It is a functional block diagram of the image identification device concerning one embodiment. It is a figure for demonstrating the outline | summary of the flow of a process by a 1st function part and a 2nd function part from the viewpoint of the data processed.

  FIG. 1 is a functional block diagram of an image identification device 1 according to an embodiment. The image identification device 1 includes a first function unit 100, a second function unit 200, and a third function unit 300. The first function unit 100 includes a mapping unit 2, a calculation unit 3, a matching unit 4, a mapping coefficient storage unit 20, and a local feature amount storage unit 40. The second function unit 200 includes an identification target storage unit 10, a mapping calculation unit 11, a preceding stage mapping unit 21, and a preceding stage calculation unit 31. The third functional unit 300 includes an imaging unit 301, a processing unit 302, and a pre-processing unit 303.

  FIG. 2 is a diagram showing an outline of the flow of processing performed in the present invention from the viewpoint of each data that is sequentially processed. In particular, the processing in the first functional unit 100 and the second functional unit 200 is shown. Has been. (Note that the third function unit 300 performs additional processing on the premise of the processing of FIG. 2 as described later.) In FIG. 2, the input image F is shown from the input image F shown on the upper side. A series of flows until F is obtained as an identification result FO [i] shown on the lower side indicating whether F corresponds to the identification object i (i = 1, 2,..., M) is shown.

  In the present invention, generally, identification results can be obtained for each of a plurality of M objects. Therefore, the description will be made assuming that there are a plurality of identification objects, but a single identification object is handled as M = 1. May be.

  The arrow line L11 drawn between the data in FIG. 2 corresponds to the processing performed in each part of FIG. 1, and the data indicated at the root of the arrow is used as an input, and is indicated at the tip of the arrow. Data is obtained as output. Hereinafter, an outline of the processing of each unit in FIG. 1 will be described with reference to each data and arrow shown in FIG.

  The first functional unit 100 receives the input image F and determines whether the input image corresponds to each identification object i (i = 1, 2,..., M) as an identification result FO [i] (i = 1, 2, ..., M) respectively. At this time, each unit functions as follows. The input image F may be received from an imaging unit such as a digital camera, or may be received as a pre-recorded image or the like via a network or the like.

The mapping unit 2 is responsible for the processing of arrows L2-1 and L2-2 in FIG. 2, and is stored in the mapping coefficient storage unit 20 for the input image F in order to identify the identification target i (in the verification unit 4). The mapping by the mapping coefficient C [i] of the identification target i is applied to obtain the mapped pixel signal F [i] _[reduced] (image constituted by the mapped pixel signal) of the input image F. Here, as will be described later in the description of the second functional unit 200, the mapping coefficient C [i] is calculated in advance to reduce the correlation between the signals in the image O [i] of the identification target i. Are stored in the mapping coefficient storage unit 20.

The calculation unit 3 is responsible for the processing of the arrow L3 in FIG. 2, and the local feature amount f ₁ specialized for identifying the identification target i from the mapped pixel signal F [i] _[reduction] obtained by the mapping unit 2. _[i] , f _{2 [i]} , ..., f _{n [i]} are calculated. As the local feature amount, a predetermined one of various known features such as SIFT, SURF, Ferns, and the like may be used.

The collation unit 4 is responsible for the processing of arrows L4-1 and L4-2 in FIG. 2, and local feature quantities f _{1 [i]} and f ₂ specialized for identifying the identification target i calculated by the calculation unit 3 are used. _[i] , ..., f _{n [i]} and local feature values o _{1 [i]} , o _{2 [i]} , ..., of the identification target i stored in the local feature value storage unit 40 o _{m [i]} is collated, and a discrimination result FO [i] indicating whether or not there is an identification target i in the input image F is obtained by threshold determination for the degree of coincidence. The identification result FO [i] may be given as a numerical value of the degree of coincidence before the threshold determination.

Here, the local feature amount o _{1 [i]} , o _{2 [i]} , ..., o _{m [i]} of the identification target _i is the identification target as described later in the description of the second functional unit 200. In the image O [i] of i, it is calculated in advance from the mapping pixel signal O [i] _[reduction] to which the mapping coefficient C [i] for reducing the signal is applied, and is stored in the local feature amount storage unit 40. Yes.

Note that the matching unit 4 determines the local feature amount f _{1 [i]} , f _{2 [i]} , ..., f _{n [i]} specialized for the identification target i and the local feature amount o ₁ of the identification target i. _[i] , o _{2 [i]} , ..., o _{m [i]} are collated to find the degree of match, using a well-known method in object recognition using local features it can. For example, for each local feature f _j (j = 1 [i], 2 [i], ..., n [i]), the local where the distance d (f _j , o _k ) is the shortest while selecting a point feature quantity o _k as a corresponding point, by combining the robust estimation of such RANSAC, it is possible to determine the degree of matching as a whole.

The second function unit 200 is the identification target i stored in the mapping coefficient storage unit 20 that is required when the first function unit 100 identifies the input image F as the identification target i as described above. Mapping coefficient C [i] and local feature quantity o _{1 [i]} , o _{2 [i]} , ..., o _{m [i] of} identification object i stored in local feature quantity storage unit 40 In addition, the first function unit 100 performs processing that is obtained in advance before the processing is performed. Note that “previous stage” in the name of the functional part represents that the process is performed in advance, and is distinguished from the functional part responsible for the same process in the first functional unit 100 (however, the processing target is different). Because.

  The identification target storage unit 10 stores an image O [i] of each identification target i (i = 1, 2,..., M). The image O [i] is prepared in advance by manual work or the like.

  The mapping calculation unit 11 is responsible for the processing of the arrow L11 in FIG. 2, acquires the image O [i] of the identification target i from the identification target storage unit 10, and reduces redundancy between signals of the image O [i]. The mapping coefficient C [i] to be calculated is calculated. The mapping coefficient C [i] calculated for each identification object i is stored in the mapping coefficient storage unit 20 and is also passed to the preceding mapping unit 21.

The preceding mapping unit 21 is responsible for the processing of arrows L21-1 and L21-2 in FIG. 2, acquires the image O [i] of the identification target i from the identification target storage unit 10, and calculates the mapping calculated by the mapping calculation unit 11 The coefficient C [i] is applied to obtain the mapped pixel signal O [i] _[reduction] . The details of the calculation by the mapping calculation unit 11 and the application by the preceding-stage mapping unit 21 will be described later.

Front calculating unit 31 is responsible for processing the arrow L31 in FIG. 2, mapping pixel signals O obtained by the pre-stage mapping unit 21 per identification target i [i] _[reduced] from the local feature amount o _{1 [i],} o _{2 [i]} , ..., o _{m [i]} is calculated. Local feature values o _{1 [i]} , o _{2 [i]} ,..., O _{m [i]} calculated for the identification target i are stored in the local feature value storage unit 40. Here, the upstream calculation unit 31 calculates the same type of local feature as the calculation unit 3.

As described above, the outline of the first functional unit 100 and the second functional unit 200 has been described. In the present invention, high-precision identification is possible based on the following principle. That is, in the present invention, the mapping coefficient C [i] for reducing the correlation between the signals is obtained from the image O [i] of the identification object i, and both the input image F and the image O [i] of the identification object i are obtained. Then, it is converted into a mapping pixel signal to which a mapping by the common mapping coefficient C [i] is applied. That is, common conversion is applied to the color space in both images. In FIG. 2, F [i] _[reduction] and O [i] _[reduction] are grayed out to indicate that the common conversion has been performed.

  The color space to be converted is calculated from the image O [i], and the local feature amount is remarkably generated for the identification target i, and the difference between the local feature amounts is emphasized. Therefore, it is possible to identify with high accuracy whether the input image F corresponds to the identification target i by collating the local feature amounts in the color space specialized for identification of the identification target i.

  Next, the third functional unit 300 will be described. The third functional unit 300 can be omitted, but when used, the third functional unit 300 achieves the following additional effect in the identification process.

  That is, when processing such as white balance and exposure correction is performed on the input image F in advance, more accurate identification is possible. The outline of the flow of processing at this time is as follows. That is, after the third function unit 300 applies in advance the same processing as that previously applied to the input image F to the image O [i] of the identification target i, the first function unit 100 and The second function unit 200 executes the same processing as described above. In FIG. 1, a flow when a similar processing process is applied in advance to the image O [i] of the identification target i is indicated by a dotted arrow. Each part of the third functional unit 300 is as follows.

  The imaging unit 301 is configured as a camera or the like, acquires an input image F (captured image F) in a state where no processing is performed, and passes it to the processing unit 302. The processing unit 302 performs processing such as white balance and exposure correction on the unprocessed input image F. Here, for the following description, the predetermined processing and the parameters in the processing are represented by the symbol e, and the image X processed by this processing is represented as X (e). .

  Based on the notation, the processing unit 302 processes the input image F to be processed image F (e), passes the processed image F (e) to the mapping unit 2 in the first function unit 100, and processes the processed image F (e). Information e (information about what processing is applied with what parameters) is passed to the upstream processing unit 303.

  The pre-stage processing unit 303 receives the image O [i] of each identification target i from the identification target storage unit 10, and uses the processing information e acquired from the processing unit 302 to input the image O [i] into the input image F. The same processing as that performed for the identification target i is performed, and the processed image O [i] (e) of the identification target i is not included. And it passes to the preceding stage mapping unit 21.

Under the above processing of the third function unit 300, the first function unit 100 and the second function unit 200 were processed instead of the input image F and the image O [i] of the identification target i, respectively. By performing processing on the images F (e) and O [i] (e), high-precision identification reflecting the processing process e is possible. That is, the local feature amounts calculated in advance for the identification target i and stored in the local feature amount storage unit 40 are also o _{1 [i] (e)} , o _{2 [i] (e)} ,..., O _{m [i ] (e)} , it is given in a form reflecting the processing e common to the input image F (e), so that it is possible to identify with high accuracy.

  In addition, when the imaging unit 301 performs imaging in real time, and the processing unit 302 determines the optimal processing for each time t image and performs the processing e (t) that changes with time, the mapping coefficient and the local feature amount Cannot be stored as fixed values (fixed values in the predetermined processing e) in the mapping coefficient storage unit 20 and the local feature amount storage unit 40.

That is, also in the second function unit 200, the mapping calculation unit 11 and the preceding mapping unit 21 perform processing in real time by reflecting the time change e (t), and the mapping coefficient C [i] referred to by the mapping unit 2 And the local feature quantity o _{1 [i] (e)} , o _{2 [i] (e)} , ..., o _{m [i] (e) of the} identification target i referred to by the matching unit 4 is expressed as e It is necessary to have a value that reflects the effect of (t) in real time. (If the processing process e is fixed, only the first functional unit 100 may perform real-time processing.)

Alternatively, in order not to generate a load on the second functional unit 200 to be reflected in real time, the following may be performed. That is, the mapping coefficient C [i] (e) and local feature values o _{1 [i] (e)} , o _{2 [i} when all the assumed image processing e is applied to the identification target image O [i]. _{] (e)} , ..., o _{m [i] (e)} are all calculated in advance by the second function unit 200 and stored in the first function unit 100, and the corresponding mapping coefficient C [ i] (e) may be applied to the input image F (e).

In the case of the above real-time processing, the first function unit 100 acquires the image processing processing information e applied at each time t, and the corresponding mapping coefficient C [i] (e) and the local feature amount o _{1 [i] (e)} , o _{2 [i] (e)} , ..., o _{m [i] (e)} are selected from the stored information and used. The flow of obtaining the information e is not shown in FIG.

  Similarly, taking into account changes in appearance due to changes in time zones, weathers, etc., for each identification target i, identification target image O [i] ( c) is prepared, and each mapping coefficient and local feature amount are retained. The image identification device 1 is obtained separately by manual input or the like as specifying the parameter c, and switches the mapping coefficient and the local feature to be applied in accordance with time information and weather information in the input image F. The identification performance in 4 can be improved.

  For example, an image O [i] (sunny) photographed in fine weather and an image O [i] [cloudy] photographed in cloudy weather are prepared as an input image for a certain identification target i. By separately acquiring information about whether shooting was performed on a fine day or cloudy day for F, when identifying the identification target i, it is possible to perform high-precision identification in consideration of the weather at the time of shooting. Similarly, for example, it is possible to distinguish between daytime, evening, and night as a time zone, thereby enabling highly accurate identification. The imaging conditions may be specified by a plurality of items such as weather and time zone.

  Hereinafter, calculation of the mapping coefficient C [i] of the identification target image O [i] of the identification target i by the mapping calculation unit 11, and the mapping by the preceding mapping unit 21 of the mapping by the calculated mapping coefficient C [i] Details of application to O [i] and application to the input image F by the mapping unit 2 will be described together with related items in other functional units.

  The specific calculation procedure of the mapping coefficient C [i] first separates the O [i] pixel of the image of the identification target i into m signals. The type and number of signals to be separated are not limited, but as an example, it is possible to use a signal in a color space such as an RGB signal, a Y UV signal, or a Y CbCr signal. Basically, it may be separated into all signals in the configuration from the beginning in the image O [i] of the identification target i. For example, if it is composed of RGB signals, it may be separated into the three signals.

Here, there are N pixels (the number of pixel positions in space) of O [j] (identified as a subscript i used separately) in the image of the identification target j, and an RGB signal ( m = 3), R _i , G _i , B _i (i = 1, 2, ..., N) are divided into m = 3 signals in each of the N pixels. ), First, the matrix P is expressed by the following (Equation 1) on the assumption that the signal is separated.

Subsequently, as in (Expression 2), the product of the matrix P and its transposed matrix P ^t is calculated as a matrix A. Here, the superscript t represents a transposition operation on the matrix.

  At this time, the matrix A is a symmetric matrix (a symmetric matrix having a size of m × m) that is the number m of signal elements in which both the number of rows and columns are separated. Therefore, the matrix A can be decomposed into matrix products as in (Equation 3).

Here, U represents an m × m orthogonal matrix, and Σ represents an m × m diagonal matrix in which singular values σ _i (1 ≦ i ≦ rank A) of the matrix A are arranged in descending order. Singular value sigma _i is the square of the eigenvalue lambda _i of A ^t A. Specific procedures degrades as the equation (3) determines the eigenvalues of A ^t A First, by calculating the singular value, determining the sigma. Next, the orthogonal matrices U and V are obtained from the definitions using U ^t U = I and V ^t V = I (I is a unit matrix), and first, the following (Equation 4) is obtained.

Accordingly, the column vector v _i of V as shown by (Equation 5) is obtained as the eigenvector corresponding to the eigenvalue sigma _i ² of A ^t A.

  In this way, the orthogonal matrix V obtained by the mapping calculation unit 11 is used as the mapping coefficient C [i], and the pixels constituting the image O [i] of the identification target i by the preceding mapping unit 21 as shown in (Equation 6) below. The mapping pixel signal P ′ applied to P and used for calculating the local feature amount by the pre-stage calculation unit 31 can be obtained. Here, regarding the position when the mapped pixel signal P ′ is configured as an image, each element in the signal P ′ may be arranged at a position in the original image O [i] of the element in the corresponding signal P.

  Note that the above (Equation 6) has been described as the process in which the pre-stage calculation unit 31 obtains the mapped pixel signal, but the calculation unit 3 also targets the input image F instead of the image O [i] of the identification target i, A mapped pixel signal of the input image F is obtained by exactly the same processing.

  Further, when only a part of the P ′ mapped pixel signal is used in the upstream mapping unit 21, only the necessary mapping may be applied. The part includes a spatial part, a temporal part (when the processing frame rate is lowered), a part of the number of signals, or any combination thereof, and is set in advance. Hereinafter, a first example and a second example will be described as an example in which only a part is used.

  As a first example, when a part of the number of signals is used and the number of signals is 1, a mapped pixel signal is obtained by the following (Equation 7).

  In addition, when using two or more as the number of signals, in both the first function unit 100 and the second function unit 200, the mapped pixel signal in each signal is obtained by the above (Equation 7), and further, the local in each signal After the feature amount is obtained, when the identification is performed by the collation unit 4, the identification result in each signal is weighted at a predetermined ratio and the final identification result is obtained.

Here, the predetermined ratio for weighting can be set based on the information amount of each signal. For example, when signals are calculated by principal component analysis as in (Expression 1) to (Expression 6), the information amount is given by the eigenvalue σ _i ² . Even when the calculation is performed by another method other than the principal component analysis, the information amount corresponding to each method may be used.

  As a second example, when using a spatial part in the image O [i], after obtaining P from only the pixel in the part rather than from the entire image, (Equation 1) to (Equation 6) ) Should be applied. In the case of using the spatial part, it is necessary to give in advance as additional information which part is used in the image O [i].

  In order to obtain the mapping coefficient C [i] as a coefficient for identifying the identification target i with higher accuracy, the region of the identification target i in the image O [i] It is preferable to use it. In this case, in order to specify the region of the identification target i from the entire image O [i], it may be specified in advance by a known object extraction technique and / or manual operation.

  Hereinafter, supplementary items (1) to (9) in the present invention will be described.

  (1) Since the color space in which the input image F is configured and the original color space of the image O [i] of the identification target i need to be common, the image O [i] is common to such Prepared in advance in a color space. For example, if the input image F is an RGB image, the image O [i] is also prepared as an RGB image.

  (2) The image F (e) obtained by performing the processing process e on the input image F. The second function unit 200 also performs the process e on the image O [i] of the identification target i to obtain O [i] ( In the embodiment designated as e), the image O [i] that is initially prepared is prepared as an unprocessed one.

(3) From the mapped pixel signal O [i] _[reduction] of the image O [i] of the identification target i in the former calculation unit 31, the local feature quantities o _{1 [i]} , o _{2 [i]} , ..., o _When calculating _{m [i]} and storing it in the local feature quantity storage unit 40, only the local feature quantity caused by the identification target i is selected, and the local feature quantity calculated from an unrelated background is excluded. It may be. In this case, information on the area occupied by the identification target in the image may be prepared in advance, and only the local feature amount calculated from the area may be selected using the information.

  In the example in which the mapping coefficient C [i] is calculated only from the pixels in the region occupied by the identification target i, an effect of automatically eliminating the local feature amount calculated from the background or the like can be obtained.

  (4) When there are a large number of identification objects i (i = 1, 2,..., M), mapping unit 2 calculates mapping coefficients C [i] for all identification objects i with respect to input image F in mapping unit 2. In order to reduce the applied load, the following may be performed. That is, clustering is performed on a set of all mapping coefficients C [i] (i = 1, 2, ..., M) by a well-known method, and similar items are grouped into one class (one group). At the same time, a quantized mapping coefficient (one mapping coefficient representing each class) is obtained for each class.

  In this way, on the first function unit 100 side, for all identification targets i belonging to the class, the mapping unit 2 obtains only one mapping pixel signal for each class of mapping coefficient that represents the class, and calculates in the same manner. Since only one local feature value common to each class needs to be obtained in part 3, the load can be reduced.

  On the second function unit 200 side, it is necessary to perform clustering after obtaining the mapping coefficient C [i] in the mapping calculation unit 11. The pre-mapping unit 21 and the pre-calculation unit 31 process the image O [i] of each identification target i using a mapping coefficient that is representative of the corresponding class, and local features under the representative mapping coefficient The amount is stored in the local feature amount storage unit 40.

  (5) When obtaining an identification result for each of a plurality of identification objects i (i = 1, 2,..., M), each of the matching degrees in each object i may be given as a numerical value, It may be possible to give the one with the largest numerical value of the matching degree as an object of the identification result, and for each target i, it is determined whether or not it is the target i by threshold determination with respect to the numerical value of the matching degree You may make it give a result.

  (6) In (Equation 1) to (Equation 7), etc., the example in which principal component analysis by singular value decomposition is used in the calculation of mapping coefficient C [i] has been described. May be. Alternatively, mapping coefficients obtained by a plurality of preset calculation methods are prepared, and when each is applied, a mapping coefficient having an excellent matching degree in the matching unit 4 is selected, and the result of the mapping coefficient is finalized. It can also be used as a typical output.

  That is, the degree of coincidence may be obtained by each calculation method, and a result that maximizes the degree of coincidence may be adopted for each identification target.

  (7) In the local feature quantity stored in the local feature quantity storage unit 40, the reference coordinate information when the image O [i] is photographed in a predetermined posture is further incorporated manually, etc. Appropriately matched identification target i applies a marker position / orientation calculation process (calculates a planar projection transformation matrix) well known in the AR (augmented reality) field, etc., so that the position / orientation is estimated. Also good.

  (8) The present invention can be similarly implemented even if the image O [i] of the identification object i is not prepared in advance and is prepared and input as an unknown one. In this case, the unknown input image F and the unknown image O [i] are input, and the output includes whether the objects included in the images match or are numerically matched. Will be obtained.

  (9) The present invention may be provided as a program that, when read and executed by a computer, causes the computer to function as part or all of the image identification device 1 of FIG.

  DESCRIPTION OF SYMBOLS 1 ... Image identification apparatus, 100 ... 1st function part, 200 ... 2nd function part, 300 ... 3rd function part, 2 ... Mapping part, 3 ... Calculation part, 4 ... Collation part, 20 ... Mapping coefficient memory | storage part, 40 DESCRIPTION OF SYMBOLS ... Local feature-value memory | storage part, 10 ... Identification object memory | storage part, 11 ... Mapping calculation part, 21 ... Pre-stage mapping part, 31 ... Pre-stage calculation part, 301 ... Imaging part, 302 ... Processing part, 303 ... Pre-stage processing part

Claims (11)

An image identification device for identifying whether an input image corresponds to an identification target,
A mapping unit configured to apply a mapping by a mapping coefficient to the input image to obtain a mapped pixel signal;
A calculation unit for calculating a local feature amount of the mapped pixel signal;
A collation unit that identifies whether the input image corresponds to the identification target by collating the calculated local feature quantity with the local feature quantity of the identification target;
The applied mapping coefficient is a mapping coefficient obtained by reducing the correlation between the signals from the image to be identified,
The image identification apparatus according to claim 1, wherein the local feature amount of the identification target is calculated in a mapping pixel signal obtained by applying a mapping based on the obtained mapping coefficient to the identification target image.   The image identification apparatus according to claim 1, wherein the mapping coefficient is obtained by principal component analysis or independent component analysis. The mapping coefficient is determined by a plurality of predetermined calculation methods,
The mapping unit obtains a mapped pixel signal for each calculation method,
The calculation unit calculates a local feature amount for each calculation method,
3. The image identification apparatus according to claim 1, wherein the collation unit obtains a collation degree at the time of identification for each calculation method, and employs a result of a calculation method that maximizes the collation degree.   4. The image identification device according to claim 1, wherein the mapping coefficient is obtained from a part of the image to be identified.   The image identification apparatus according to claim 4, wherein the part is an area occupied by the identification target in the identification target image. The matching unit performs matching using a local feature amount in each signal of the mapped pixel signal to obtain a matching degree, and weights the matching degree in each signal based on the information amount of each signal. , Perform the identification,
The mapping coefficient is obtained by principal component analysis or independent component analysis,
The image identification apparatus according to claim 1, wherein the information amount is an information amount at the time of the analysis. There are a plurality of identification objects,
The mapping coefficient is calculated for each identification object and then clustered to determine a representative mapping coefficient for each class,
The local feature amount of the identification target is calculated by a representative mapping coefficient in the corresponding class,
The mapping unit applies a representative mapping coefficient of each class to obtain a mapping pixel signal of each class,
The calculation unit calculates a local feature amount of each class,
The said collation part collates the local feature-value calculated by the said calculation part in the class corresponding in the said identification object, and the local feature-value of the said identification target, The one of Claim 1 thru | or 6 characterized by the above-mentioned. The image identification device described in 1. Means for acquiring information of processing performed in advance on the input image;
The image identification apparatus according to claim 1, further comprising: a unit that applies in advance an image processing based on the acquired processing information to the image to be identified.   The image identification apparatus according to claim 8, wherein the obtaining unit and the applying unit perform processing in real time in conjunction with an input image obtained in real time. Means for obtaining an imaging condition in the input image;
The image to be identified is prepared for each imaging condition,
10. The image identification apparatus according to claim 1, wherein the collation unit performs the collation by using a local feature amount of the identification target corresponding to the acquired imaging condition.   An image identification program for causing a computer to function as the image identification apparatus according to claim 1.