JP5791373B2 - Feature point position determination device, feature point position determination method and program - Google Patents

Description

  The present invention relates to a feature point position determination device, a feature point position determination method, and a program for determining a plurality of feature point positions in an image.

  In personal recognition using face image data (hereinafter referred to as face recognition), the determination of the position of a facial organ or a characteristic part corresponding to the face organ (hereinafter referred to as a feature point) is an important task, and regulates recognition performance. There are many cases. In addition, the determination of the position of feature points with high accuracy requires a high processing load and may limit the time of the entire recognition process.

  Therefore, for example, Patent Document 1 discloses a method for speeding up processing by reducing the number of feature points extracted in the current frame by using the recognition result of the previous frame when an individual is recognized from moving image data. Has been. A technique for determining the positions of a plurality of feature points is to select feature point candidates based on a feature point candidate detector that extracts position candidates for each feature point and the feature point arrangement specific to the target object (for example, a face). In many cases, a correction processor is used for correction. For example, Non-Patent Document 1 discloses a method of determining a plurality of facial organ feature positions according to geometric constraints based on statistics.

JP 2009-75999 A

Beumer, G.M .; Tao, Q .; Bazen, A.M .; Veldhuis, R.N.J. "A landmark paper in face recognition" Automatic Face and Gesture Recognition, 2006. FGR 2006. 7th International Conference, pp. 73-78

  However, the technique disclosed in Patent Document 1 is severely degraded in recognition performance by reducing the number of feature points to be extracted. In addition, the method disclosed in Non-Patent Document 1 is a method for efficiently correcting the position of a feature point using a partial space. However, when the extraction accuracy of a feature point position candidate is low, it cannot be corrected effectively. There is.

  In view of the above-mentioned problems, the present invention realizes appropriate feature point position correction even when the accuracy of feature point position candidates detected for speeding up the processing is low, and determines a desired feature point position. The purpose is to be able to.

The feature point position determination device of the present invention is a feature point position determination device that determines the positions of a plurality of feature points from image data, a candidate determination unit that obtains position candidates of the plurality of feature points, and the candidate determination unit Determination means for determining the reliability of the position candidate of the feature point obtained by the above, condition determination means for determining a correction condition having a higher binding force as the reliability is lower based on the reliability, and the plurality of features characterized by chromatic and correcting means for correcting, based the position candidate of a point on the correction condition.

  According to the present invention, the position of a feature point can be appropriately determined even when the detection accuracy of the feature point position candidate is low.

It is a flowchart explaining a feature point position determination process and a face recognition process. It is a block diagram which shows the structural example of the image recognition apparatus which concerns on embodiment. It is a figure explaining the example of a clipping process of face image data. It is a figure which shows the 15 feature-point position relevant to a facial organ. It is a figure explaining the example of a process of step S103 of FIG. It is a figure explaining the example of template matching. It is a figure explaining the example of a correction process. It is a figure explaining the processing content of step S108 of FIG. It is a flowchart which shows an example of the procedure of the operation | movement in 2nd Embodiment. It is a figure explaining the flow of step S902 of FIG.

(First embodiment)
The operation of the first embodiment of the present invention will be described below with reference to FIGS.
FIG. 2 is a block diagram illustrating a configuration example of the image recognition apparatus 200 according to the present embodiment. First, the image recognition apparatus 200 extracts a region including a face from image data. Hereinafter, the extracted area is referred to as face image data. A plurality of feature point positions (here, positions of features related to facial organs) are determined from the obtained face image data, and a recognition function for identifying an individual based on the feature point positions is provided.

  In FIG. 2, reference numeral 201 denotes an image input unit. The image input unit 201 includes an optical device, a photoelectric conversion device, a driver circuit that controls a sensor, an AD converter, a signal processing circuit that controls various image corrections, a frame buffer, and the like. Reference numeral 202 denotes a preprocessing unit that executes various types of preprocessing in order to effectively perform various types of subsequent processing. Specifically, image data conversion such as color conversion processing and contrast correction processing is performed on the image data acquired by the image input unit 201 by hardware.

  The face image data cutout processing unit 203 performs face detection processing on the image data corrected by the preprocessing unit 202. Various conventionally proposed methods can be applied to the face detection method. The face image data cutout processing unit 203 normalizes and cuts out face image data to a predetermined size for each detected face.

  FIG. 3 is a diagram illustrating an example of face image data cutout processing. The face area 32 is determined from the image 31 corrected by the preprocessing unit 202, and the upright face image data 33 normalized to a predetermined size is cut out. Thus, the size of the upright face image data 33 is constant regardless of the face. The cut face image data is stored in a RAM (Random Access Memory) 209 via a DMAC (Direct Memory Access Controller) 205. Hereinafter, the position of the feature point is defined as the coordinate of the feature point in the erect face image data 33, and is expressed by a coordinate system (x coordinate, y coordinate) with the upper left end of the erect face image data 33 as the origin. And

  Reference numeral 207 denotes a CPU (Central Processing Unit), which executes main processing according to the present embodiment and controls the overall operation of the image recognition apparatus 200. A bridge 204 provides a bus bridge function between the image bus 210 and the CPU bus 206. Reference numeral 208 denotes a ROM (Read Only Memory), which stores instructions that define the operation of the CPU 207. A RAM 209 is a work memory necessary for the operation of the CPU 207. The CPU 207 executes recognition processing on the face image data stored in the RAM 209.

  FIG. 1 is a flowchart for explaining feature point position determination processing and face recognition processing according to the present embodiment. The flowchart shows the operation of the CPU 207. In this embodiment, a feature point position candidate determination method is selected according to the operation mode, and a suitable feature point position correction process is executed according to the selected feature point position candidate determination method.

  First, an operation mode is determined in step S101. The operation mode here is a normal mode or a tracking mode. In the normal mode, face recognition is performed with priority given to recognition accuracy over processing speed. On the other hand, in the tracking mode, high-speed recognition processing is executed while allowing deterioration in accuracy. Whether the mode is the normal mode or the tracking mode is determined based on whether or not a recognition target person (person registered as a recognition target) has been recognized in the previous frame. For example, the operation mode is determined with reference to the recognition result of the previous frame stored in the RAM 209.

  Next, in step S102, information for selecting a feature point position candidate determination method is acquired according to the determination result in step S101. Table 1 below shows an example of table information for selecting a feature point position candidate determination method for each feature.

  In the normal mode, a normal mode table is selected in step S102, and in the tracking mode, a tracking mode table is selected. Contents 1 and 2 described in the table of Table 1 correspond to the first feature point position candidate determination method and the second feature point position candidate determination method, respectively.

  FIG. 4 is a diagram for explaining an example of the position of the feature point determined here. Fifteen feature point positions related to the facial organs 401 to 415 are determined. In step S102, feature point position candidates corresponding to the feature points are selected with reference to the selected table. For example, in the normal mode, the first feature point position candidate determination method is selected for all feature points with reference to the table shown in Table 1. On the other hand, in the tracking mode, a different feature point position candidate determination method is selected for each feature point. For example, at the feature point 401, the process proceeds to step S103, and at the feature point 403, the process proceeds to step S104.

  In step S103, feature point position candidates are determined with high accuracy. Here, feature point position candidates are determined using a highly accurate method that requires processing time. On the other hand, in the second feature point position candidate determination method, feature point position candidates are determined by a method that prioritizes processing time over accuracy. The contents of the table shown in Table 1 are determined in advance in consideration of the desired processing speed and performance.

  FIG. 5 is a diagram illustrating an example of the process in step S103. In the present embodiment, feature point position candidates are determined by CNN (Convolutional Neural Networks). FIG. 5 shows a configuration in the case where two feature point position candidates are determined for the sake of explanation. The CNN is configured by hierarchical feature extraction processing.

  FIG. 5 shows an example of a two-layer CNN in which the number of features in the first layer 506 is 3 and the number of features in the second layer 510 is 2. Reference numeral 501 denotes face image data, which corresponds to the above-described upright face image data 33. Reference numerals 503 a to 503 c denote characteristic surfaces of the first hierarchy 506. The feature plane is an image data plane that stores a result of calculation while scanning data of the previous layer with a predetermined feature extraction filter (cumulative sum of convolution calculations and nonlinear processing). Since the feature plane is a detection result for raster-scanned image data, the detection result is also represented by a plane. 503a to 503c are calculated using different feature extraction filters with reference to the face image data 501. 503a to 503c are generated by two-dimensional convolution filter calculation corresponding to 504a to 504c and non-linear conversion of the calculation result, respectively. Reference numeral 502 denotes a reference image area necessary for the convolution calculation. For example, a convolution filter operation with a kernel size (length in the horizontal direction and height in the vertical direction) of 11 × 11 is processed by a product-sum operation of Expression (1) as shown below.

  Here, input (x, y) is a reference pixel value at coordinates (x, y), and output (x, y) is a calculation result at coordinates (x, y). weight (column, row) is a weight coefficient at coordinates (x + column, y + row), and columnSize = 11 and rowSize = 11 are filter kernel sizes (number of filter taps).

  Reference numerals 504a to 504c denote convolution filter kernels having different coefficients. Also, the size of the convolution kernel varies depending on the feature plane. In the CNN operation, a product-sum operation is repeated while scanning a plurality of filter kernels in units of pixels, and a final product-sum result is nonlinearly transformed to generate a feature plane. A sigmoid function or the like is applied to the nonlinear transformation. When calculating 503a, since the number of connections with the previous layer is 1, there is one convolution filter kernel 504a. On the other hand, when calculating the feature plane 507a and the feature plane 507b, since the number of connections with the previous layer is 3, the calculation results of the three convolution filters corresponding to 508a to 508c and 508d to 508e, respectively, are cumulatively added. That is, the feature plane 507a is obtained by accumulating all the outputs of the convolution filters 509a to 509c and finally performing a nonlinear conversion process. Reference numerals 505a to 505c denote reference image areas necessary for the convolution calculation of the second hierarchy 510, respectively. The coefficient of the convolution filter is determined in advance by learning.

  In step S103, the center of gravity of the values of the feature planes 507a and 507b, which are CNN calculation results, is used as the feature point position candidate coordinates. In the case of the present embodiment, a network that can determine 15 feature point position candidates is actually configured using a CNN having 15 feature numbers in two layers. A necessary feature plane of the second layer is calculated according to the result of step S102, and feature point position candidates are determined with high accuracy. CNN calculation is known as a powerful feature extraction method, but it is a method that requires such a high processing load.

  On the other hand, in step S104, feature point position candidates are determined by a method with a low processing load. In this embodiment, the feature point position is determined by simple template matching. FIG. 6 is a diagram for explaining an example of template matching. Here, an example in which the nose 409 in FIG. 4 is simply detected will be described. First, the similarity between the template 64 corresponding to the nose prepared in advance and the pixel data in the predetermined search area 62 in the face image data 61 is calculated. The similarity is calculated by normalized correlation calculation between the template and the pixel value. The similarity is calculated while scanning 63 the template 64 within a predetermined area, and a nose position candidate is determined based on the coordinates where the peak value of the similarity appears, the center of gravity in the search area 62 of the similarity, and the like. The template matching calculation can determine feature point position candidates with a smaller processing load than the CNN calculation described in the first feature point position candidate determination method. However, the robustness with respect to variations in individual differences and face orientation is low and the overall accuracy is low.

  Next, in step S105, it is determined whether or not all feature point position candidate determination processes have been completed. With the above processing, feature point position candidates are determined by different methods for each feature point according to the operation mode. The CPU 207 stores the determined feature point position candidates in the RAM 209.

  Through the above steps S101 to S105, all feature points are determined with high accuracy by CNN calculation in the normal mode. On the other hand, in the tracking mode, feature points 401, 402, 405, 406, 412, and 415 are determined with high accuracy by CNN calculation, and other feature points are determined at high speed using template matching. In this way, only some feature points having high importance in the subsequent processing are calculated with high accuracy.

  Next, in step S106, the reliability of the feature point position candidate is determined. In this embodiment, the reliability is determined according to the operation mode. That is, it is determined that the normal mode has high reliability and the tracking mode has low reliability.

  In step S107, the projection dimension of the eigenvector used for the geometric correction process described later is determined. Here, the projection dimension is selected according to the reliability determined in step S106. When it is determined that the reliability of the feature point candidate position is high, a high dimension number is selected, and when it is determined that the reliability of the feature point candidate position is low, a low dimension is selected. In the present embodiment, the correction condition in step S108 described later is changed by changing the projection dimension.

  In step S108, geometric correction processing is executed on the 15 feature point position candidates obtained in step S103 or step S104, and the final feature point position is determined. FIG. 7 is a diagram illustrating an example of correction processing. 402a is a feature point that characterizes the corner of the eye but is erroneously determined as the position of the end of the eyebrows. In step S108, the position is corrected by statistical processing based on the arrangement relationship of human face features. In the example illustrated in FIG. 7, 402a is corrected to the position 402b based on the arrangement relationship of the plurality of feature point candidate positions.

FIG. 8 is a diagram for explaining the processing content of step S108. In step S801, each feature point position candidate coordinate is simply connected to generate one vector data. In this embodiment, 30-dimensional feature vector data V is generated from 15 feature point position coordinates. A data string obtained by simply connecting the position coordinate data (x _i , y _i ) (i: feature point numbers 1 to 15) of each feature point is referred to as feature point vector data V (element v _j : j = 1 to 30). To do. The feature point numbers 1 to 15 correspond to the feature points 401 to 415 in the embodiment. Therefore, for example, the elements v ₁ and v ₂ of the feature point vector data correspond to the x coordinate value and the y coordinate value of the feature point 401, respectively. The feature vector data V is defined by the following equation (2). Hereinafter, T represents transposition.
V = (v ₁ , v ₂ , v ₃ ,... V ₂ × _f ) ^T ... Equation (2)

  Here, f indicates the number of feature points.

In steps S802 and S803, a projection vector is calculated using an average vector A807 and a projection matrix E808, respectively. The projection vector P is calculated by the following equations (3) to (5) using the vector data obtained by subtracting the average vector A from the feature point vector data V and the projection matrix E. Note that the projection matrix E and the average vector A are previously calculated by principal component analysis using feature point vector data (learning feature vector data) for a large number of face images. The learning feature vector data is vector data generated by connecting all the correct feature point position coordinates of the face image in the same manner.
P = E ^T (VA) (3)
A = (a ₁ , a ₂ , a ₃ ,... A ₂ × _f ) ^T (4)
E = (u ₁ , u ₂ ,... U _p ) (5)

u ₁ , u ₂ ,..., u _p are 2 × f-dimensional orthonormal vectors (eigenvectors) obtained by principal component analysis. In the case of this embodiment, it is a 30-dimensional vector. p is the dimension of the projection vector, and a different value is used according to the reliability of the feature point candidate position. That is, the projection dimension used here is the value set in the projection dimension determination processing step S107. Specifically, it is set to 8 in the normal operation mode, that is, when the accuracy of the feature point candidate is high, and is set to 6 in the tracking mode, that is, when the accuracy of the feature candidate is low. In this case, the projection matrix E is a matrix in which eight vectors having large corresponding eigenvalues are selected from the orthogonal vectors obtained by the principal component analysis, and only six of the large eigenvalues are used in the tracking mode. Become. Assume that the projection matrix E and the average vector A are stored in advance in the ROM 208, the RAM 209, or the like.

In steps S802 and S803, the 2 × f-dimensional feature point vector is reduced to a p-dimensional projection vector by the calculations of equations (3) to (5). That is, it projects onto a p-dimensional subspace. In steps S804 and S805, the original feature point vector dimension data (that is, the coordinate position) is restored from the projection vector P. The restoration vector V ′ is calculated by the following equation (6) using the projection matrix E and the average vector A.
V ′ = EP + A (6)

  In step S806, the corrected coordinate data is extracted from the back-projected restored vector data V ′. Through the processing from step S801 to step 806 described above, the statistical outlier can be corrected by projecting the vector data obtained by concatenating all the feature point position data onto the subspace whose dimension has been reduced and then performing reverse projection. That is, outliers (incorrect detection) that cannot be expressed in the projected subspace are corrected. As a result, the geometrical arrangement relationship is corrected based on the arrangement relationship of each feature point, and the erroneous detection as indicated by 401a in FIG. 6 is corrected.

  Here, when the projection dimension number is high, the degree of freedom of the space projected with the eigenvector is high, and when the projection dimension number is low, the degree of freedom of the space projected with the eigenvector is low. That is, when the projection dimension number is high, the constraint force of the geometric arrangement relationship is weak, and when the projection dimension number is low, the constraint force is increased. In this embodiment, when the accuracy of the feature point position candidate determination method is high using this feature, the position of the feature point position determination method is corrected by selecting a high projection dimension and correcting the position of the feature point with weak constraints. If is low, a low projection dimension is selected and the position of the feature point is corrected with strong constraints. For example, when the projection dimension is set to 0, the result of the feature point position correction processing is an average vector A807, which is equivalent to not using all the results of the feature point position candidate determination method (not trusting all). In the tracking mode, since the number of feature points determined at a high speed in step S104 is large, it is determined that the reliability is low, and a strong constraint is given.

  Returning to the description of FIG. 1, in step S109, recognition processing is executed based on the position of the feature point obtained in step S108. For recognition processing, various conventionally proposed methods may be applied. For example, a plurality of local regions are cut out based on the determined feature point positions, and dimensionally compressed by orthogonal transformation or the like. The dimension-compressed data is used as a feature vector (feature amount). Similarly, the degree of similarity with the registrant is calculated by a correlation operation with the calculated feature vector of the registered person. The feature vector of the registered person is stored in the RAM 209 or the like prior to recognition.

  A plurality of feature vectors are calculated according to the positions of feature points. For example, it is calculated from a plurality of local areas including eyes, nose and mouth. The final similarity is calculated by integrating the calculated correlation values of the plurality of feature vectors. Whether or not the user is a registrant is determined by thresholding the final similarity. The number and type of local areas to be used are the same even when the operation modes are different.

  As described above, in the present embodiment, processing is performed by changing the detection accuracy of the feature point candidate position according to the operation mode, and the projection dimension of the geometric constraint processing is selected according to the detection accuracy (reliability of the feature point candidate position). In step S110, the recognition result is recorded in the ROM 208.

  According to the present embodiment, even when high-speed processing is performed while allowing accuracy degradation such as the tracking mode, the feature point position can be appropriately determined while suppressing performance degradation. That is, it is possible to realize an appropriate correction according to the feature point position candidate determination method by a very simple method of changing the number of projection dimensions.

(Second Embodiment)
In the first embodiment, the case where the reliability is determined according to the operation mode has been described. In this embodiment, the reliability is determined from target face image data. FIG. 9 is a flowchart illustrating an example of an operation procedure according to the present embodiment.

  In step S901, feature point position candidates are determined. Here, similarly to step S103 of the first embodiment, a plurality of feature point position candidates are obtained by CNN calculation. In step S902, the feature point reliability is determined. Here, the reliability of the feature point position candidate determination process is estimated from the contrast and resolution of the target face image. For example, if the contrast is low, the contrast is sharp, the resolution is low, and the image is blurry, the reliability of step S901 is determined to be low.

  FIG. 10A illustrates the flow of step S902 based on the contrast histogram. In process 1001, a histogram of pixel values of the erect face image data 33 is created. Then, the reliability of the feature position candidate is determined by a discriminator such as SVM (Support Vector Machine) using the histogram obtained in the processing 1002 as a vector. The SVM learns the relationship between the histogram shape and the reliability of the feature point position candidates in advance.

  FIG. 10B illustrates the flow of step S902 based on the resolution. In step 1003, edge extraction filter processing is executed for the erect face image data 33, and in step 1004, the sum of edge amounts is calculated. In step 1005, the total value is compared with a predetermined threshold value to predict the resolution. When the resolution is high, it is determined that the reliability of the feature point position candidate is high.

  FIG. 10C is a diagram illustrating the flow of step S902 using the image data itself. First, erecting face image data 33 is subsampled to a predetermined size in process 1006, and the data subsampled in process 1007 is determined as a vector by SVM. The SVM learns the relationship between the subsampled image data and the reliability of the feature point position candidates in advance.

  Returning to the description of FIG. 9, in step S903, the projection dimension is set in accordance with the reliability determined in step S902. A high projection dimension is set when the feature point position reliability is high, and a low dimension is set when the reliability is low. Steps S904 to S906 are the same as steps S108 to S110 of FIG. 1 described in the first embodiment.

  As described above, in the present embodiment, the reliability of the feature point position candidate determination process is predicted by analyzing the target image, and the projection dimension is set based on the prediction. That is, when the reliability is predicted to be low, the geometric correction processing constraint force is increased, and when the reliability is predicted to be high, the geometric correction processing constraint force is decreased. Accordingly, it is possible to determine an appropriate feature point position according to various face images to be processed.

(Other embodiments)
In the first embodiment, the case where the operation mode (normal mode / tracking mode) is selected according to the recognition result of the previous frame has been described. However, the present invention is not limited to such a case. For example, the high accuracy mode and the high speed mode are switched depending on the number of faces in the target image. The present invention can be applied to various cases such as a user setting a high-precision mode and a high-speed mode through a predetermined user interface. For example, when setting via a user interface, step S101 determines an operation mode based on information specified by the user.

  In the second embodiment, the case where the reliability of the feature point position candidate is predicted by image analysis on the target face image has been described. However, other methods may be used. For example, a method of determining the reliability based on the information (intermediate result) obtained in step S901 may be used. In other words, when the feature point candidates are determined by the CNN shown in the embodiment, the reliability can be determined using the output value of the CNN. When the output value of CNN is high, the possibility of the target feature point is high. Therefore, it is possible to predict by a method such as determining reliability based on the sum of CNN output values of all feature point position candidates. is there. Even when the feature point position candidate determination process is realized by simple template matching, it can be similarly determined using the sum of correlation values or the like.

  In the above-described embodiment, the case where the projection onto the partial space (step S803) and the reverse projection (step S804) are processed independently has been described. It is also possible to do. Since each of these operations is a linear matrix operation, a transformation matrix that integrates the projection operation and the reverse projection operation can be generated and processed in advance.

  Furthermore, in the above-described embodiment, the case where a specific person is recognized from the face image in the image has been described, but the present invention is not limited to this. The present invention can be used for various image recognition apparatuses that recognize or detect a predetermined object based on the arrangement of feature points. Furthermore, the present invention can be applied to a method for recognizing states such as the posture and facial expression of the target object.

  In the above-described embodiment, the case where the position of the feature point is a position on the two-dimensional coordinate has been described. However, the present invention is not limited to this, and in the case of processing based on three-dimensional data, the feature point is applied to the three-dimensional coordinate. Also good. In this case, feature vector data V is created by concatenating three coordinate values. Furthermore, in the embodiment, the case where the present invention is applied to an image recognition apparatus has been described. However, the present invention is not limited to this, and the present invention can also be used for an apparatus that corrects and deforms an image using the coordinates of determined feature points.

  The present invention can also be realized by executing the following processing. That is, software (program) that realizes the functions of the above-described embodiments is supplied to a system or apparatus via a network or various storage media, and a computer (or CPU, MPU, or the like) of the system or apparatus reads the program. It is a process to be executed.

207 CPU
208 ROM
209 RAM

Claims (8)

A feature point position determination device that determines the positions of a plurality of feature points from image data,
Candidate determination means for obtaining position candidates of the plurality of feature points;
Determination means for determining the reliability of the position candidate of the feature point obtained by the candidate determination means;
Based on the reliability, condition determination means for determining a correction condition with a higher binding force as the reliability is lower;
Wherein the plurality of feature point location determination apparatus characterized by chromatic and correcting means for correcting, based the position candidate of the feature points on the correction condition. 2. The feature point position determining apparatus according to claim 1, wherein the correction condition of the correction means is a geometrical constraint force of an arrangement relation of a plurality of feature points. The correcting means is obtained by connecting the coordinate data of a plurality of feature point position candidates and inputting them as vector data, projecting means for projecting the vector data onto a subspace of a predetermined number of dimensions, and the projecting means. Reverse projection means for reversely projecting the projected vector, and means for extracting coordinate data from the result of the reverse projection means,
2. The feature point position determination apparatus according to claim 1, wherein the correction condition of the correction means is the number of dimensions. Means for determining the operation mode;
2. The feature point position determination apparatus according to claim 1, wherein the determination unit determines the reliability according to a result of the unit that determines the operation mode. The determination means includes image analysis means for analyzing image data,
2. The feature point position determination apparatus according to claim 1, wherein the reliability is determined based on a result of the image analysis means.   The feature point position determination apparatus according to claim 1, wherein the determination unit determines the reliability based on information obtained by the candidate determination unit. A feature point position determination method for determining the positions of a plurality of feature points from image data,
A candidate determination step for obtaining position candidates of the plurality of feature points;
A determination step of determining the reliability of the position candidate of the feature point obtained in the candidate determination step;
Based on the reliability, a condition determination step for determining a correction condition with a higher binding force as the reliability is lower;
Feature point position determination method characterized by chromatic and a correction step of correcting, based the position candidate of the plurality of feature points on the correction condition.   The program for making a computer perform each process of the feature point position determination method of Claim 7.

Images