JP2006163614A - Image processing device and method - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide an image processing device capable of maintaining detection identification accuracy, and capable of detecting a target object with a small processing amount. <P>SOLUTION: This image processing device comprises a reference database 11 having a plurality of reference parameters 12 which are combinations of a reference vector including a target object reference vector as an image vector of a target object and a non-target object reference vector as an image vector of a non-target object and a weight coefficient of the reference vector; a partial image cutting out part 3 for cutting out a partial image from an input image 2; a feature quantity vector calculating part 4 for calculating a feature quantity vector of the partial image cut out; and a feature quantity vector evaluating part 5 for detecting the target object from the partial image with the usage of the feature quantity vector and the reference parameter. The feature quantity vector evaluating part includes an image identification part 6 for determining including/non-including of the target object of the partial image, based on the calculated result, and detection processing is completed when the image identification part determines that the target object is not included. <P>COPYRIGHT: (C)2006,JPO&NCIPI

Description

  The present invention relates to an image processing apparatus and an image processing method for detecting a target object from an image.

  Conventionally, in the utilization of surveillance cameras such as security, it is required to detect a desired object from an image. For example, “face” is detected as an object in an image. As such an apparatus or method for detecting a specific object from an image, an apparatus or a method for identifying by comparing a parameter having a vector between a face and a face other than the face (hereinafter, non-face) and an input image has been proposed. (For example, refer to Patent Documents 1 and 2).

  In the description of Patent Document 1, first, a reference parameter learned by a boosting technique is input. Next, a weighted linear sum of output values of the image identification unit is calculated as an evaluation value. Further, the target object is distinguished from the other objects according to the calculated evaluation value, and the final detection result of the target object is output. It has been experimentally confirmed that this boosting method can generate a reference parameter having high discrimination performance.

  FIG. 17 is a block diagram of an image processing apparatus in the prior art, and represents the image processing apparatus described in Patent Document 1.

  First, an arbitrary part of the input image is cut out as a partial image by the partial image cutout unit 1801. The partial image is cut out by scanning various size windows starting from a predetermined position of the image while sequentially shifting an appropriate pixel (for example, one pixel) to the right side or the lower side. Next, the feature amount vector calculation unit 1802 calculates a feature amount vector from the partial image. The feature vector evaluation unit 1803 calculates an evaluation value using the reference parameter and the feature vector of the partial image. At this time, the reference parameter is a parameter having a target object and a non-target object vector learned in advance by the boosting technique, and an evaluation value is calculated using these parameters, so that the target object or the non-target object can be calculated. Which is similar is determined.

  With this process, for example, when a face is a target object, a face and a non-face are identified. Note that it has been experimentally confirmed that the boosting method for generating the reference parameters has high discrimination performance.

  Next, FIG. 18 is a block diagram of an image processing apparatus in the prior art, and represents the image processing apparatus disclosed in Patent Document 2.

  First, the partial image cutout unit 1901 cuts out a partial image from the input image in the same manner as in FIG. Next, the feature quantity vector evaluation unit 1902 calculates a feature quantity vector of the clipped partial image. Further, the feature vector evaluation unit 1902 calculates an evaluation value using the reference parameter and the feature vector of the partial image. It is determined based on the calculated evaluation value whether the partial image includes the target object or not. Here, the feature quantity vector evaluation unit 1902 includes a plurality of image identification units 1903, and the image identification unit 1903 calculates an evaluation value. Here, the plurality of image identification units 1903 calculate evaluation values using different reference parameters. As shown in FIG. 18, the image identification units 1903 are connected in series and execute processing in order. In the apparatus of FIG. 18, the process ends when any of the image identification units 1903 connected in series determines that the partial image does not include the target object. After the process is completed, no other image identification unit is used. For this reason, high-speed processing is realized.

  However, in the method of Patent Document 1, since the feature vector is evaluated using all the learning parameters generated by the boosting technique, the amount of processing required for the evaluation becomes enormous and real-time processing cannot be performed. Had.

Further, in the method of Patent Document 2, a plurality of image identification units to which different reference parameters are given evaluates feature quantity vectors. Here, different reference parameters differ in performance for identifying a target object. In the processing using the processing in the plurality of image identification units to which the reference parameters having different identification performances are given in this way, the identification may be performed in order from the reference parameters having the lowest identification performance. For this reason, there has been a problem that even if high-speed processing is possible, detection processing with high identification accuracy cannot be performed.
US Patent Application Publication No. 2003/0108244 US Patent Application Publication No. 2002/0102024

  SUMMARY OF THE INVENTION Accordingly, an object of the present invention is to provide an image processing apparatus and an image processing method that can maintain identification accuracy for detecting a desired target object from an image and perform target object detection with a smaller processing amount than in the past. To do.

  An image processing apparatus according to a first invention is an image processing apparatus that detects a target object from an input image, and includes a target object reference vector that is an image vector of the target object and a non-target object reference that is an image vector of the non-target object. Reference database having a plurality of reference parameters that are combinations of reference vectors including vectors and weighting coefficients of reference vectors, a partial image cutout unit that cuts out a partial image from an input image, and features of the partial image cut out by the image cutout unit A feature vector calculation unit that calculates a quantity vector, and a feature vector evaluation unit that performs a detection process of detecting a target object from a partial image using the feature vector and the reference parameter. Inclusion / non-wrapping of the target object of the partial image based on the calculation result calculated using the quantity vector and the reference parameter A determining image identification section, an image identification portion, and terminates the detection process when it is determined that the non-inclusion of the target object.

  With this configuration, when the partial image cut out from the input image does not include the desired target object, the detection process can be completed, so the processing amount can be reduced and high-speed processing can be realized. It becomes.

  The image processing apparatus according to the second invention ends the detection process when there are a plurality of image identification units and at least one of the plurality of image identification units determines that the target object is not included.

  As a result, the image processing in the image identification unit can be sequentially executed with different reference parameters, and the detection processing can be terminated when any of the image identification units is determined not to be included during execution. Processing can be realized.

  In the image processing apparatus according to the third invention, the image identification unit calculates a correlation value for calculating a target object correlation value between the feature quantity vector and the target object reference data, and a correlation value between the feature quantity vector and the non-target object reference data. A target object correlation value calculated by the correlation value calculation unit, an evaluation value calculation unit for calculating an evaluation value that is a product of the non-target object correlation value and the weighting factor, and inclusion of the target object of the partial image according to the evaluation value / The determination part which performs non-inclusion determination is provided.

  With this configuration, it is possible to perform inclusion / non-inclusion determination of the target object of the partial image cut out using the closeness of the reference parameters of the target object and the non-target object, so that highly accurate identification can be performed. .

  According to a fourth aspect of the present invention, there is provided the image processing apparatus, wherein the determination unit includes an intermediate determination unit that repeatedly performs inclusion / non-inclusion determination using different reference parameters when the calculation processing in the evaluation value calculation unit is a predetermined number of times or less And a final determination unit that performs inclusion / non-inclusion determination when the calculation processing in the evaluation calculation unit is a predetermined number of times or more.

  With this configuration, when an identification process based on an arbitrary number of times is executed, the process is forcibly terminated at that time, so that the detection process can be performed with a processing amount equal to or less than a predetermined number according to the specification. . Thereby, high-speed processing becomes possible.

  In the image processing apparatus according to the fifth invention, when the feature vector evaluation unit determines that at least one of the plurality of image determination units includes the target object, the detection processing is continued using different reference parameters.

  With this configuration, even when it is determined that the target object is included, the identification process can be continued using still another reference parameter, so that the accuracy of detection of the target object can be further improved.

  In the image processing device according to the sixth aspect, the feature vector evaluation unit uses a plurality of different reference parameters in descending order of the weighting coefficient.

  Since it is possible to execute the inclusion / non-inclusion identification process of the target object of the partial image using the reference parameters in descending order of the weighting factors, it is possible to execute the identification process sequentially from the reference parameter that easily identifies the target object. As a result, the detection process can be performed efficiently and the necessity of continuing the detection can be determined at an early stage, so that high-speed processing can be realized.

  In the image processing device according to the seventh aspect, the feature vector evaluation unit uses the target object reference data in the ascending order of the number of dimensions of the non-target object reference data regarding the non-target object.

  With this configuration, even when evaluation is performed using a multidimensional feature quantity vector, detection processing can be executed in order of higher discriminating power, so that efficient detection processing can be realized.

  In the image processing apparatus according to the eighth invention, the reference parameter is generated by ensemble learning from the target object reference vector and the non-target object reference vector.

  With this configuration, the reference parameter used for the identification process is generated in a learning manner, so that it is possible to use a reference parameter with high identification ability.

  In the image processing apparatus according to the ninth aspect of the invention, the ensemble learning uses a boosting technique.

  With this configuration, ensemble learning for generating reference parameters can be easily and appropriately executed.

  In the image processing apparatus according to the tenth aspect, the feature vector is calculated by a filter process for calculating at least one of an edge feature and a multi-resolution feature (hereinafter referred to as “wavelet feature”) by Haar wavelet transform.

  With this configuration, a feature vector can be generated for each level of frequency components, and identification processing can be executed in that order, so that highly efficient detection can be executed. Furthermore, detection corresponding to the case of an input image that has been wavelet transformed in advance can be easily performed.

  In the image processing apparatus according to the eleventh aspect, the wavelet feature is obtained by octave division.

  With this configuration, it is possible to efficiently execute identification processing from an image corresponding to the wavelet transform level.

  In the image processing device according to the twelfth aspect, the feature vector evaluation unit sequentially uses high-level components by octave division.

  Since the identification process is executed from the component corresponding to the level generated by the wavelet transform, the identification process can be executed efficiently. Furthermore, since the amount of processing can be reduced, the processing speed can be increased.

  In the image processing apparatus according to the thirteenth invention, the input image is at least one of encoded and decoded in a wavelet compatible format.

  With this configuration, it is possible to execute identification processing using reference parameters based on wavelet transformation without performing wavelet transformation on the input image again.

  In the image processing apparatus according to the fourteenth aspect, the wavelet compatible format corresponds to the JPEG2000 format.

  With this configuration, it is possible to easily cope with the case where identification processing is performed on an input image corresponding to the JPEG2000 format captured by a digital camera or the like, so that the target property using electronic devices having various imaging functions Can be detected.

  An image processing method according to a fifteenth aspect of the invention is an image processing method for detecting a target object from an input image, wherein the target object reference vector is an image vector of the target object and the non-target object reference is an image vector of a non-target object. A parameter setting step for setting a plurality of parameters, which is a combination of a reference vector including a vector and a weighting factor of the reference vector, a partial image cutting step for cutting a partial image from the input image, and a feature vector of the cut partial image A feature amount vector calculating step for calculating, and a feature amount vector evaluating step for performing a detection process for detecting a target object from the partial image using the feature amount vector and the parameter. The feature amount vector evaluating step includes a feature amount vector and a parameter. Inclusion of target objects of partial images based on the calculation results calculated using And an image determining step of determining non-contained, a plurality of the image determining step, and ends the detection process when it is determined that the non-inclusion of the target object.

  With this configuration, when the partial image cut out from the input image does not include the desired target object, the detection process can be completed, so the processing amount can be reduced and high-speed processing can be realized. It becomes.

  In the image processing method according to the sixteenth invention, the image determination step calculates a correlation value for calculating a correlation value between the target object correlation value between the feature vector and the target object reference data, and a correlation value between the feature vector and the non-target object reference data. And an evaluation value calculating step of calculating an evaluation value that is a product of the step and the correlation value and the weighting coefficient.

  With this configuration, the image processing in the image identification unit is sequentially executed with different reference parameters, and the detection process can be ended when it is determined that the target object is not included in any of the image identification units during execution. High-speed processing can be realized.

  According to the present invention, the detection accuracy is not deteriorated by performing the detection process in order from the reference parameter having the high identification accuracy. Furthermore, since the evaluation is terminated when it is determined as a non-target object, speeding up can be realized at the same time.

(Embodiment 1)
In the first embodiment of the present invention, an image processing apparatus and an image processing method for detecting a face from an image will be described as an example of object detection. Note that the target object is not limited to a face, and may be another desired object.

  FIG. 1 is a block diagram of an image processing apparatus according to Embodiment 1 of the present invention.

  First, the configuration of the image processing apparatus 1 will be described.

  The image processing apparatus 1 includes a partial image cutout unit 3 that cuts out a partial image from the input image 2, a feature amount vector calculation unit 4 that calculates a feature amount vector of the cut out partial image, and a feature amount vector and a reference parameter 12. A feature vector evaluation unit 5 for determining inclusion / non-inclusion of an object is provided. Furthermore, a reference database 11 including a reference parameter 12 is provided. The reference parameter 12 includes a target object reference vector 13, a non-target object reference vector 14, and a weighting factor 15. The feature vector evaluation unit 5 includes a plurality of image identification units 6, and ends the detection process when any one of the image identification units 6 determines that the partial image does not include the target object. . Further, the image identification unit 6 includes a correlation value calculation unit 7 that calculates a correlation value between the feature quantity vector and the target object reference vector 13 and a correlation value between the feature quantity vector and the non-target object reference vector 14. Further, a similarity calculation unit 8 that calculates the similarity between the feature quantity vector and the target object from the correlation value calculated by the correlation value calculation unit 7, and an evaluation value that is a product of the calculated similarity and the weight coefficient 15 An evaluation value calculation unit for calculating, and a determination unit for determining whether the target object of the partial image is included or not from the evaluation value are provided.

  Finally, the detection result is output.

  Next, the detailed configuration and operation of each unit will be described.

  First, the partial image cutout unit 3 will be described.

  The partial image cutout unit 3 cuts out the input image 2 to an arbitrary size. The arbitrary size may be a predetermined size in units of pixels, for example, and the size is also arbitrary. Alternatively, the upper left vertex of the input image 2 may be sequentially moved to the right as the start point, moved to the lower left when reaching the right end, and finally cut out in the order of the lower right as the end point. The partial image cutout unit 3 realizes image cutout by, for example, applying a window having a predetermined size to the input image 2.

  The partial image cutout unit 3 may cut out partial images from all of the input images 2 in advance and continuously perform target object detection processing. Alternatively, the partial image cutout unit 3 cuts out one partial image and extracts the target object from the partial image. This processing may be performed to cut out the next partial image.

  Next, the feature quantity vector calculation unit 4 will be described.

  The feature vector calculation unit 4 calculates a feature vector from the partial image. The feature amount vector is a one-dimensional arrangement of feature amounts extracted from the partial images. The feature amount used in image processing includes image texture information itself, an average of pixel values of images of a plurality of detection target objects, and the like. These are high-dimensional features that use color (luminance in the case of a grayscale image) information in an image and position information of an object as feature amounts, and may be easily affected by the strength of a light source.

  On the other hand, edge information may be used as a feature amount instead of color information and luminance information. The edge information is obtained by applying a differential filter in the x direction and the y direction, and represents the contour of the object, shape information, and the like. Furthermore, a value obtained by normalizing edge information with intensity may be used as the feature amount. This value is obtained by dividing the edge information in the x direction and the y direction by the intensity, and is a vector representing the direction component of the edge feature. Therefore, this value is referred to as an edge normal direction vector in this specification. Since the edge normal direction vector normalizes the edge with the edge intensity that is the brightness information, the edge normal direction vector has a characteristic that is robust against the brightness change.

  Here, calculation of the edge normal direction vector will be described. First, a Sobel filter is applied to the partial image. The Sobel filter is in the x direction

  For the y direction

  It is a filter which has these. By these Sobel filters, an xy component that is an intensity in the x direction and an intensity in the y direction is obtained. This xy component is normalized by intensity.

  Next, an edge image in which an edge normal direction vector is calculated for each pixel of the partial image is obtained. This edge image is enlarged or reduced to a predetermined window size (36 × 36 in this embodiment). Thereafter, the pixel values of the scaled edge image are arranged in one column to be a feature vector. At this time, the window size is preferably the same size as the reference parameter of the reference database 11. This is because the feature vector evaluation unit 5 can be easily processed.

  As described above, a feature vector based on edge information is calculated.

  Next, the feature vector evaluation unit 5 will be described.

  The feature vector evaluation unit 5 includes a plurality of image identification units 6. The image identification unit 6 includes a correlation value calculation unit 7, a similarity calculation unit 8, an evaluation value calculation unit 9, and a determination unit 10. A plurality of image identification units 6 are provided, and each image identification unit 6 reads different reference parameters 12 from the reference database 11. That is, each image identification unit 6 detects inclusion / non-inclusion of the target object of the partial image according to different reference parameters. The reference parameter 12 includes a target object (face in the present embodiment) reference vector 13, a non-target object reference vector 14, and a weighting factor 15 to be detected. By including each of the target object reference vector 13 and the non-target object reference vector 14, it is easily determined whether the partial image is similar to the target object or similar to the asymmetric object.

  The correlation value calculation unit 7 calculates a correlation value between the feature quantity vector using the target object reference vector 13 and the non-target object reference vector 14 included in the read reference parameter. The correlation value calculation is realized by convolution calculation of both vectors. As the correlation value, both the correlation value with the target object reference vector 13 and the correlation value with the non-target object reference vector 14 are calculated. The calculation result of the correlation value calculation unit 7 is output to the similarity calculation unit 8.

  The similarity calculation unit 8 calculates the similarity to the target object or non-target object based on the input correlation value. Since the correlation value is calculated for both the target object and the non-target object, the degree of the correlation value is directly calculated as the similarity. For example, if the correlation value with the target object is high, the similarity to the target object is high, and if the correlation value with the non-target object is high, the similarity to the non-target object is high. The calculated similarity is output to the evaluation value calculation unit 9.

  Since the similarity is worthy of replacement of the correlation value, the evaluation value may be calculated directly from the correlation value without calculating the similarity.

  The evaluation value calculation unit 9 calculates an evaluation value that is a product of the weight coefficient 15 and the similarity. Here, the weighting coefficient 15 is a coefficient representing the importance of the reference parameter 12 used in the image identification unit 6. That is, the degree of identification accuracy of the used reference parameter 12 is represented. The higher the weighting coefficient, the higher the identification accuracy, indicating that the reference parameter is easy to identify the target object and the non-target object. In the determination of inclusion / non-inclusion of the target object in the image identification unit 6, there is an advantage that determination with higher identification accuracy is possible by multiplying not only the correlation value with the reference vector but also the weight coefficient. In addition, when a reference parameter with a low weight coefficient, that is, a reference parameter with low identification accuracy is used, the evaluation value is reduced by multiplying the weight coefficient, and inaccurate identification is not performed. There are also benefits.

  The evaluation value calculated by the evaluation value calculation unit 9 is output to the determination unit 10. The determination unit 10 determines inclusion / non-inclusion of the target object of the partial image from the input evaluation value. For example, if the evaluation value is large in a direction similar to the target object, it is determined that the partial image targeted for detection processing includes the target object. If the evaluation value is large in a direction similar to the non-target object, the target object is not included. Judge as.

  Here, the image identification unit 6 given the different reference parameters 12 sequentially evaluates the feature quantity vectors of the clipped partial images. At this time, the detection process ends when any of the plurality of image identification units 6 determines that the target object is not included. This is because when the target object is a face, if it is determined that the partial image is a non-face, there is no need to continue the process. Conversely, if it is determined that the face is the target object, the process is continued because there is a possibility that a non-face is included. At this time, if it is determined that the face is non-face by executing the processing in order from the reference parameter having the high discrimination ability, the processing can be terminated because the face is not included. After the detection process is completed, the partial image cutout unit 3 cuts out the partial image at the next position, the detection process is continued for the new partial image, the process for the entire input image 2 is completed, and the handling object is detected. It becomes possible to do.

  As described above, the detection processing is performed in order from the reference parameter 12 having the highest weighting coefficient in the plurality of image identification units 6, so that high-speed processing can be realized without reducing the identification accuracy in detection.

  Here, the configuration in which a plurality of image identification units 6 are provided has been described. However, the configuration is such that the image identification unit 6 is a single unit, and different reference parameters 12 are read into the single image identification unit 6 one after another and the corresponding processing is performed. But you can.

  Next, another form of the feature vector evaluation unit 5 will be described.

  FIG. 2 is an internal configuration diagram of the feature vector evaluation unit 5 according to Embodiment 1 of the present invention.

  The feature vector evaluation unit 5 includes an evaluation value calculation unit 21, an intermediate determination unit 22, a final determination unit 23, a reference parameter holding unit 24, and a control parameter holding unit 25. The evaluation value calculation unit 21 calculates an evaluation value of the input feature vector. The intermediate determination unit 22 and the final determination unit 23 determine whether or not the face is based on the evaluation value. The control parameter holding unit 25 holds a determination parameter indicating whether the evaluation value is input to the intermediate determination unit 22 or the final determination unit 23, and the reference parameter storage unit 24 includes the intermediate determination unit 602 and the final determination unit 603. Parameters necessary for determining whether or not the face is a face and a threshold value used for determination processing in each determination unit.

  FIG. 3 is a process flowchart of the feature vector evaluation unit 5 according to Embodiment 1 of the present invention.

  The operation of the feature vector evaluation unit 5 shown in FIG. 2 will be described.

  First, in step 701a, the evaluation value calculation unit 21 reads a reference parameter from the reference parameter holding unit 24 and calculates an evaluation value. The reference parameter stores M weighting factors αj corresponding to the reference vector Fj of the target object, the reference vectors NFj and Fj of the non-face, and NFj (1 ≦ j ≦ M). The weight coefficient αj represents the discrimination performance of the learning results Fj and NFj, and the discrimination performance is higher as the value of αj is larger. In the reference parameter holding unit, reference parameters are stored in the order in which they are generated regardless of the values of Fj, NFj, and αj.

  Next, in step 701 b, the evaluation value calculation unit 21 reads the determination parameter p from the control parameter holding unit 25 and selects either the intermediate determination unit 22 or the final determination unit 23. In the present embodiment, the determination parameter p has an initial value (0 ≦ p ≦ M) in advance. Here, M is the number of generated reference parameters. Next, in step 701c, every time the intermediate determination unit 22 performs processing, p decreases by the value “1”. When p is 1 or more, the intermediate determination unit 22 performs processing, and when p is 0, the final determination unit 23 performs processing. That is, the determination parameter confirms the progress of the detection process for the number of reference parameters, and when the determination process is performed for the number of reference parameters, the process is performed regardless of whether the target object is determined. Is terminated.

  In step 701c, the intermediate determination unit 22 performs determination processing. In step 701d, as a determination process, the calculated evaluation value is compared with a threshold value. As a result of the comparison, if it is equal to or less than the threshold value, the partial image to be processed is determined as a non-face, and if it is larger than the threshold value, it is determined as a face. When the partial image is determined to be a face, the calculation process of the evaluation value calculation unit 21 is executed again. The evaluation value calculation unit 21 newly reads a reference parameter that has not been applied to the feature under evaluation from the reference parameter holding unit 24, and recalculates the evaluation value. On the other hand, if it is determined as a non-face, the determination result is output and the process ends.

  Finally, when the determination parameter p becomes “0”, in step 701e, the final determination unit 23 outputs a final determination result as to whether or not the face is a face.

  As described above, the end of the process can be determined at high speed by dividing the intermediate determination unit 22 and the final determination unit 23.

  Next, the processing of the evaluation value calculation unit 21, the intermediate determination unit 22, and the final determination unit 23 will be described with calculation formulas used.

  First, the evaluation value calculation unit 21 calculates an evaluation value according to (Equation 3).

  Here, Xn is a feature amount vector, h (Fj, NFj, Xn) is an evaluation function of the evaluation value calculation unit 21, and a reference parameter (face reference) input from the feature amount vector Xn and the reference parameter holding unit 24 The vector Fj and the non-face reference vector NFj) are used as arguments. αj is a weighting factor corresponding to the reference parameters Fj and NFj. The calculation unit 21 reads m reference parameters and their weight coefficients from the reference parameter holding unit 24, and calculates an evaluation value by using these weighted linear sums. Note that the number m of reference parameters to be read and the reading order are set in advance, and are appropriately set depending on the target object or image.

  This evaluation function h (Fj, NFj, Xn) is defined by (Equation 4).

  (Equation 4) calculates normalized correlation values of the feature vector Xn, the face reference vector Fj, and the non-face reference vector NFj, respectively, and compares the correlation values of Fj and NFj. As a result of comparison, the value “1” is returned if the feature vector Xn has a large correlation value with Fj, and the value “−1” is returned if the correlation value with NFj is large. As a result, when it is determined that the target object is included, the value “1” is returned, and when it is determined that the target object is not included, the value “−1” is returned.

  Xn (−), Fj (−), and NFj (−) are average vectors in which average values of the vectors Xn, Fj, and NFj are stored.

  Next, the intermediate determination unit 22 and the final determination unit 23 determine the evaluation value H (Xn) calculated in (Equation 3) as a target object or a non-face based on the determination threshold th (p). To do. The determination threshold th (p) represents a determination threshold when the determination parameter is p. When the evaluation value H (Xn) is greater than th (p), the face is determined. Otherwise, the face is determined as a non-face. The determination threshold th (p) is input from the reference parameter holding unit 24.

  The determination threshold th (p) is set so that the face reference vector is not included below th (p). By using this determination threshold value th (p) in the intermediate determination unit 22, when it is determined that the evaluation value is equal to or less than the determination threshold value, this partial image can be clearly determined as a non-face. If it is determined that the face is non-face, no further identification processing is required, and the processing can be terminated as it is. Conversely, if it is equal to or greater than the determination threshold, it is determined as a face, and the determination process is continued using a new reference parameter.

  The final determination unit 23 executes determination processing using the last of the stored reference parameters and outputs the result. If the final determination unit 23 is used, the identification process ends regardless of whether the determination is a face or a non-face. In this case, the process proceeds to the next partial image. It should be noted that speeding up of the process can be prioritized by making the condition of the value of the determination parameter p entering the final determination process larger than the value “0”.

  Further, the reference parameter read by the evaluation value calculation unit 21 is the descending order of the weighting coefficient. The magnitude of this weighting factor represents the level of identification accuracy of the reference parameter. For this reason, by reading and determining in descending order of the weighting coefficient, it is possible to sequentially determine from the reference parameter having information that can easily identify the face and the non-face. The determination processing according to the descending order of the weighting factors enables efficient identification while maintaining the identification accuracy. For this reason, the amount of processing can be reduced, and real-time processing can be performed in the detection of the target object from the input image.

  The evaluation value calculation unit 21 and the intermediate determination unit 22 may be processed as shown in FIG.

  FIG. 4 is a process flowchart of the feature quantity vector evaluation unit in the first embodiment of the present invention, and FIG. 5 is an image of a face image and a non-face image in the first embodiment of the present invention.

  First, in step 801a, a correlation value between the feature vector Xn and the face reference vector Fj is obtained by (Equation 4). The evaluation value calculation unit 21 outputs the calculated correlation value to the intermediate determination unit 22 and the final determination unit 23. The intermediate determination unit 21 and the final determination unit 23 perform face / non-face determination based on the magnitude of the correlation value, not determination based on the evaluation value.

  As in the case of FIG. 3, the determination method compares the correlation value with the correlation threshold value ThCor (p) in step 801d, determines that the face is less than the correlation threshold value, and determines that the face is greater than or equal to the correlation threshold value. The processes performed by the intermediate determination unit 21 and the final determination unit 23 are the same as those described with reference to FIG.

  Thus, by performing face / non-face determination using only the correlation value with the face reference vector, there is an advantage that determination can be performed without being affected by the non-face reference vector. For example, as shown in FIG. 5, even if the front face in the left column is a different person, the eyes and nose and mouth are in the same position and shape so that the face reference vector converges to a certain value or level. Is possible. On the other hand, since the non-face as the background covers all but the front face, there are various shapes such as a house and a mountain. For this reason, it is difficult to converge the non-face reference vector to a certain value or level. For this reason, using a non-face reference vector for determination may cause a heavy load on the determination process. Considering such a case, performing the determination using only the correlation value with the reference vector of the face that is the target object is highly advantageous in high-speed processing or the like.

  Next, the reference database 11 will be described.

  The reference database is a database having reference parameters 12 including target object reference vectors, non-target object reference vectors, and weighting coefficients. Here, the reference parameter is generated by learning from the face and non-face images that are the target objects.

  A process for obtaining parameters by the learning process will be described.

  FIG. 6 is a flowchart showing the flow of the learning process. First, an image database is prepared. In the image database, face and non-face images used in the learning process were collected. Face images and non-face images of various forms, shapes, and sizes used for the learning process are stored in the image database. For example, the face image is an image in which only a face is cut out in advance, and the non-face image is a background image that does not include a face such as a forest or a town. Based on this image database, reference parameters are generated by a learning process.

  First, in step 1001a, feature vectors of face images and non-face images stored in the image database are calculated. Similar to the calculation of the feature amount vector of the partial image, the feature amount vector is represented by an edge normal direction vector. Edges in the x and y directions are extracted from the face and non-face images by the Sobel filter shown in (Equation 1) and (Equation 2), and divided by the edge strength to calculate the edge normal direction vector. Is done. Further, an edge image is obtained from the edge normal direction vector. The image is scaled to a certain size (36 pixels × 36 pixels in this embodiment), and the feature value vector is obtained by arranging the pixel values of the edge image in one column.

  Next, calculation of reference parameters is executed using feature vectors of faces and non-faces.

  First, the number of feature vectors is defined as Vnum. In step 1001b, teacher data Ti (1 ≦ i ≦ Vnum) is assigned to the feature vector Xi. The teacher data Ti is set to a value “+1” when assigned to a facial feature vector, and a value “−1” when assigned to a non-facial feature vector.

  Next, in step 1001c, a reference parameter is generated by ensemble learning using the feature vector Xi and the teacher vector Ti. A plurality of reference parameters suitable for detection of the target object can be generated by learning reference parameters.

  Ensemble learning is a learning method that can generate a plurality of reference parameters with slightly different discriminating powers in order to improve generalization ability. Examples of ensemble learning include bagging and boosting. FIG. 7 shows a configuration diagram of a learning unit that performs ensemble learning.

  The learning unit uses a feature vector storage unit 1101 in which feature vectors of faces and non-faces are stored, a filter unit 1102 that samples some data in the feature vector storage unit, and data output from the filter unit. A learning processing unit 1103 that calculates a reference parameter by ensemble learning, an evaluation data storage unit 1104 that stores a feature quantity vector for evaluation, and an evaluation unit 1105 that performs evaluation using the evaluation data are included.

  First, the filter unit 1102 samples the feature amount vector stored in the feature amount vector storage unit 1101. Data sampled by the filter 1102 is output to the learning processing unit 1103. The learning book rib 1103 generates a reference parameter based on the input data. At this time, an ensemble learning method such as boosting is used.

  The generated reference parameter is evaluated by the evaluation unit 1105 based on the evaluation data stored in the evaluation data unit 1104. The evaluation data is a parameter having a general-purpose value in advance, and the next new reference parameter is generated based on the comparison result with this parameter.

  A plurality of different reference parameters can be generated by the above processing procedure.

  The learning method based on ensemble learning is described in “T, G, Dietrich.“ Ensemblable methods in machine learning ”In Multiple Classifier Systems. 1st First International Works, 2000. . Therefore, further explanation is omitted here.

  Thus, M reference parameters used in (Equation 3) and (Equation 4) are generated. The larger the number M, the more accurate the identification becomes, but the calculation amount required for determining the feature vector at the time of detection increases. Therefore, the number M of reference parameters may be determined according to the purpose.

  Finally, the overall operation of the image processing apparatus and the image processing method according to Embodiment 1 will be described with reference to FIGS.

  FIG. 8 is a flowchart of the detection process in the first embodiment of the present invention, and FIG. 9 is a process diagram of the detection process in the first embodiment of the present invention.

  In FIG. 8, (X, Y) is the upper left coordinate of the detection position, S × S is the detection size, WIDTH, HEIGHT is the number of horizontal and vertical pixels of the image, and SIZEMAX is the larger value of WIDTH and HEIGHT. It is.

  First, in step 401a, the upper left coordinates (X, Y) of the detection position and the detection size S × S are initially set. In this embodiment, the upper left coordinate of the image is the origin, the origin (0, 0) is the initial detection position, and the initial detection size is 20 × 20 pixels. After the initial setting, in step 401b, the partial image cutout unit 3 cuts out partial images from (X, Y) to (X + S, Y + S) of the input image 2. Next, in step 401c, the feature quantity vector calculation unit 4 calculates a feature quantity vector from the extracted partial image. Next, in step 401d, the feature quantity vector evaluation unit 5 determines the inclusion / non-inclusion of the target object of the partial image using the feature quantity vector and the reference parameter.

  The above processing is shifted rightward by an appropriate pixel (one pixel in this embodiment) to the lower right of the input image (step 401e, step 401f), or scanned while shifting downward (step S401g, step S401h). Thus, it is determined whether or not the target object is included by sequentially cutting out partial images of S × S pixels. Further, the size S of the partial image is enlarged by an appropriate multiple (1.2 times in the present embodiment) up to the image size (320 × 240 in the present embodiment) (steps S401i and S401j). It may be processed.

  The above-described image processing apparatus and image processing method for detecting a target object may be realized by a general-purpose apparatus shown in FIG.

  FIG. 10 is a configuration diagram of an electronic apparatus including the image processing apparatus according to Embodiment 1 of the present invention.

  A CPU (Central Processing Unit) 101 executes the image processing program described in FIG. 9 stored in a ROM (Read Only Memory) 102.

  A RAM (Random Access Memory) 104 and a hard disk 105 store a reference database including reference parameters in addition to storing input images. In addition to this, intermediate data in the middle of image processing may be stored. A camera 108 is connected to the bus 103 via an interface 107, and an image captured by the camera 108 is input as an input image. A target object detection process is executed on the input image.

  The camera 108 may be either a still camera or a video camera, and a camera attached to a mobile phone can also be used.

  Moreover, you may implement | achieve with a dedicated apparatus instead of such a general purpose apparatus.

  As described above, according to the present embodiment, among the plurality of reference parameters, it is possible to perform identification in order from the reference parameter with high performance for identifying the target object and the non-target object, and the partial image is set as the non-target object. When the determination is made, the identification of the partial image can be completed at that time, so that the processing amount for the non-target object occupying most of the image can be reduced. As a result, it is possible to speed up the detection process without reducing the identification accuracy.

(Embodiment 2)
In the second embodiment of the present invention, a technique for detecting a target object at high speed and with high accuracy by using a wavelet feature obtained by Haar wavelet transform as a feature amount will be described.

  For convenience of explanation, the target object is described as a face (a face such as a person or an animal). However, the image processing apparatus and the image processing method described in the second embodiment are technologies limited to face detection. Absent.

  FIG. 11 is a block diagram of an image processing apparatus according to Embodiment 2 of the present invention, FIG. 12 is an internal block diagram of a feature vector evaluation unit according to Embodiment 2 of the present invention, and FIG. It is an operation | movement flowchart of the feature-value vector evaluation part in Embodiment 2 of invention.

  The image processing apparatus according to the second embodiment has the same overall configuration as the image processing apparatus described in the first embodiment. Here, in the configuration shown in FIG. 11, the input image 1201 and the partial image cutout unit 1202 are not different from those described in Embodiment 1, and thus the description thereof is omitted.

  First, the feature vector calculation unit 1203 will be described.

  The feature vector calculation unit 1203 calculates a feature vector from the clipped partial image. In the second embodiment, a method of using a frequency feature that is less affected by the intensity of the light source instead of the edge information of the first embodiment will be described as the feature amount.

  Here, the wavelet transform, which is one of the frequency features, can analyze and extract the original signal by expanding and contracting the size of the wavelet used for the conversion. Therefore, using wavelet transform has an advantage that both frequency information on the image by frequency transformation can be utilized in addition to the shape information of the image from the original signal.

  The feature vector calculation unit 1203 performs multi-resolution analysis by Haar wavelet transform on the partial image. Next, wavelet expansion coefficients and scaling coefficients (hereinafter referred to as wavelet features) obtained by multi-resolution analysis by Haar wavelet transform are arranged in a line. The wavelet features arranged in a line are used as a feature vector.

  Here, the wavelet transform and multiresolution analysis will be described.

  First, the wavelet will be described. The basic wavelet in the wavelet is called mother wavelet ψ (x). By this mother wavelet, a frequency component on the image is detected. That is, information on frequency components in each part on the image is obtained. Further, the mother wavelet is stretched and translated to an appropriate position to synthesize the original signal. The wavelet transform is defined by (Equation 5).

  In (Equation 5), b is a shift parameter, and a> 0 is an enlargement / reduction parameter, which is called a scale. 1 / √a is a coefficient for normalization. The inner product of ψ () and the signal f (x) is the defining formula for wavelet transform.

  Mother wavelets include Haar wavelets, Mexican hat wavelets, and Gabor wavelets. Among them, the Haar wavelet has a merit that the calculation cost can be shortened compared with other wavelets because the calculation formula is simple. Therefore, in Embodiment 2, a Haar wavelet is used.

  The Haar wavelet function is defined by (Equation 6).

  Next, multiresolution analysis will be described.

  Multi-resolution analysis is to approximate a signal with a linear combination of functions called scaling functions. Here, the accuracy of approximation of the original signal in the multi-resolution analysis is expressed as a level. By multi-resolution analysis, a signal is expressed by a linear combination of wavelet features calculated at a plurality of resolutions. By using the result of performing multi-resolution analysis on an image as a feature amount, detection evaluation can be performed using a feature amount including frequencies in a plurality of bands. Therefore, if the face area of the input image is high resolution, the correlation is high at the high resolution component, and if it is low resolution, the correlation is high at the low resolution component, and robust detection can be performed against the difference in resolution. .

  The scaling function of the Haar wavelet is defined as (Equation 7).

  Using this scaling function, an arbitrary function f (x) is defined as (Equation 8).

  Next, the translation and enlargement / reduction of the Haar scaling function are defined by (Equation 9).

  Using this φj, s, an approximation of level j is defined by (Equation 10).

  Here, sk (j) is called a scaling factor. Here, when the level is “0”, the approximation is the most accurate, and as the level value increases, the approximation becomes rough. That is, it can be said that f1 (x) lacks information compared to f0 (x). If this missing part is g1 (x), (Equation 11) is established.

  At this time, g1 (x) is called a level 1 wavelet component. From this wavelet φ, functions φj, k that have been translated and expanded / reduced are obtained as shown in (Equation 12), and when g1 (x) is expressed using φj, k, (Equation 13) is obtained. At this time, ωk (1) is called a level 1 wavelet expansion coefficient.

  As described above, the approximate function f0 (x) having the level value “0” is decomposed into the approximate function f1 (x) having the level value “1” and the function g1 (x) represented by the wavelet having the level value “1”. It is expressed as (Equation 14). When this is generalized, (Expression 15) is obtained.

  As described above, the function f0 (x) can be expressed by the sum of wavelet components having different widths, that is, different resolutions. When this is used recursively, the approximation function f0 (x) at level 0 is expressed as (Equation 16) using approximation functions at levels 1, 2,..., J.

  Next, a case where multiresolution analysis by wavelet transform is applied to an image will be described.

  A two-dimensional wavelet transform is executed on a two-dimensional signal called an image by repeatedly using a one-dimensional two-part filter bank. Specifically, first, wavelet transformation is performed in the horizontal direction of the image, 2: 1 down-sampling is performed, and a low-frequency component and a high-frequency component are separated. Next, a low-frequency component and three high-frequency components are obtained by subjecting the result to wavelet transform in the vertical direction. By repeating these operations for low frequency components, image data subjected to multi-resolution analysis can be obtained. In this way, two-dimensional filtering can be performed in a format in which horizontal and vertical processing are separated. This has the advantage that the design of the two-dimensional filter can be easily performed as a design problem of the one-dimensional filter. FIG. 14A is an image of an original image according to the second embodiment of the present invention, and FIG. 14B is an image image obtained by performing level 3 multiresolution analysis on FIG. Thus, the upper left of the image is a low frequency component, and the rest of the image is a high frequency component.

  Wavelet expansion coefficients and scaling coefficients obtained by performing multi-resolution analysis are called wavelet features. The features arranged in one column in order from the component with the highest wavelet feature level are defined as a feature vector.

  Furthermore, the wavelet feature has a feature that it can be decoded by using the encoded data to sequentially increase the resolution from the lowest resolution low resolution image to the high resolution image. This is known as spatial resolution scalability in encoding and decoding of JPEG 2000 (ISO / IEC FCD 15444-1), which is a compression encoding method for natural images. Further, when the image is decoded using the highest level component, the color and shape information of the original image at the resolution can be sufficiently expressed.

  Next, processing of the feature vector evaluation unit 1204 will be described.

  A processing flow of the feature vector evaluation unit 1204 is shown in FIG. The evaluation value calculation unit 1301 reads the reference parameter and the weight from the reference parameter holding unit 1304 and calculates an evaluation value. Here, in the present embodiment, reference parameters and weights are read as follows using the properties of wavelet features. First, evaluation is performed using only high-level components of the feature vector. As described above, the shape information of the original image is retained even with only a high-level component, so that highly accurate identification is possible. When only high-level components are used, the number of dimensions of the feature vector used for evaluation is reduced, so that the processing amount can be reduced. By performing evaluation by the intermediate determination unit 1302 based on the evaluation value thus calculated, it is possible to determine a non-face with an evaluation with a small processing amount.

  Next, evaluation is performed using components having lower levels in order. By using a component having a low level, the number of dimensions increases and the amount of information increases. Therefore, discrimination with higher accuracy is possible, and a face and a non-face can be discriminated more precisely. Thus, the overall processing amount required for face detection can be reduced.

  In the second embodiment, the evaluation value is calculated by (Equation 17).

  Here, Xn is a feature quantity vector, h (Fj, NFj, Xn, L) is an evaluation function of the evaluation value calculation unit 1301, and the feature quantity vector Xn and the face parameter input from the reference parameter holding unit 1304 The reference vector Fj, the non-face reference vector NFj, and the dimension number L of the feature vector used are used as arguments. Αj is a weight for the j-th reference parameter, and m is the number of face and non-face reference vectors and the weighting factor α corresponding to the result, which are input from the reference parameter holding unit. Here, for the sake of simplification, it is assumed that the feature quantities stored in the feature quantity vector Xn, the learning results Fj, and NFj are stored in order from the higher-level component of the wavelet feature. Therefore, by changing the number of dimensions L to be used, it is possible to specify up to which level of components to use. In the present embodiment, level 3 multi-resolution analysis is performed on a 48 × 48 image. If only the highest level component is present, the dimension number L is 12 × 12. When the next level component is used, L is 24. When using all components, × 24, L was set to 48 × 48.

  There are the following two methods for realizing the above.

  One is a method in which Fj and NFj are generated using all feature quantity vectors, and evaluation is performed using the L dimension during detection.

  Further, one is a method of generating Fj and NFj using L-dimensional components from high-level components of wavelet features in the feature quantity vector, and evaluating at the time of detection.

  In the first method, since Fj and NFj are generated using all the features obtained from the multiresolution analysis result, there is an advantage that the dimension number L can be flexibly changed at the time of detection. In the next method, since Fj and NFj are generated using the L-dimensional feature quantity vector among the feature quantity vectors obtained from the multiresolution analysis result, the L dimension is compared with the first method. It is considered that all the information obtained by learning is included.

  There are a plurality of methods for generating Fj and NFj, but an optimal method may be selected depending on the learning environment and the target image. In the second embodiment, what is learned by the second method is used. Accordingly, all m reference parameters read from the reference parameter holding unit 1204 are L-dimensional feature quantity vectors.

  By (Equation 18), an evaluation value is calculated by a weighted linear sum using m reference parameters input from the reference parameter holding unit 1204.

  The evaluation function h () is defined by (Equation 18).

  (Equation 18) calculates a normalized correlation value between the feature vector Xn, the face reference vector Fj, and the non-face reference vector NFj, and returns 1 if the correlation value with Fj is large, while NFj Is a function that returns -1 if the correlation value between and is large. Si (−), Fj (−), and NFj (−) are average vectors in which average values up to the L dimension are stored in each of Si, Fj, and NFj.

  The intermediate determination unit 1302 and the final determination unit 1303 determine whether the evaluation value H (Xn) calculated in (Equation 17) corresponds to a face or a non-face based on a determination threshold. When the evaluation value H (Xn) is larger than the determination threshold, the face is determined, and otherwise, the face is determined as a non-face. The determination threshold is input from the reference parameter holding unit.

  The detection process is performed by the method described above. In order to perform detection, the reference parameter must be stored in the reference parameter holding unit 1304 in advance.

  Here, FIG. 15 is a flowchart of reference parameter generation in Embodiment 2 of the present invention. This will be described with reference to FIG.

  The generation of the reference parameter is the same as that described in the first embodiment. However, since the detection is performed using the wavelet transform when performing the detection process, it is necessary to generate the reference parameter using the wavelet transform.

  As described in the description at the time of detection, a wavelet feature that is a result of performing multi-resolution analysis by Haar wavelet transform on a learning image is used as a feature amount. First, the face and non-face images are scaled to a certain size (36 pixels × 36 pixels in this case), and multi-resolution analysis is performed on them. Of the results obtained from the multi-resolution analysis, the feature quantity vector is obtained by arranging the components of the level necessary for learning in one column. By performing learning processing on the feature quantity vector, a reference parameter generated from the L-dimensional feature quantity vector is generated. By performing learning processing on feature quantity vectors of a plurality of dimensions (for example, 12 × 12, 24 × 24, and 48 × 48) generated by multi-resolution analysis, a multi-dimensional reference format can be generated.

  When wavelet features are obtained in the course of encoding and decoding in a wavelet-compatible format and wavelet features are obtained, the wavelet features are used as feature quantity vectors to perform processing required for the conversion work. The target object can be detected without any problem. In particular, in JPEG2000, not only encoding and decoding by wavelet multi-resolution analysis but also a function of sequentially synthesizing a plurality of resolution levels can be used. This function is characterized in that data is transferred from a server to a mobile terminal, a PC, or the like in order from a high-level component of a multiresolution analysis result. In the present embodiment, since the evaluation values can be calculated in order from the higher level components, the processing can be performed without waiting for the transfer of all the data, so that efficient processing is possible.

  Further, the image processing apparatuses shown in Embodiments 1 and 2 may be configured as shown in FIG.

  FIG. 16 is a block diagram of an image processing apparatus according to Embodiment 2 of the present invention. The image processing apparatus shown in FIG. 16 evaluates three feature quantity vectors of a feature quantity vector evaluation unit (leftward) 1704, a feature quantity vector evaluation part (front) 1705, and a feature quantity vector evaluation part (rightward) 1706. The face that is the target object is detected. This is because even the same face is different in ease of detection depending on its orientation, and therefore it is preferable to detect it using reference parameters corresponding to each face.

  The feature amount vector corresponding to the face orientation is independently evaluated by each evaluation unit. At this time, it is possible to perform face detection corresponding to the front and left / right orientation of the face by holding the reference parameters used for detection in advance as parameters that match the front and left / right orientation of the face. As a result, face detection in a plurality of directions is possible, and for example, more flexible detection processing can be performed in security or the like.

  As described above, among the features obtained by multi-resolution analysis, first, detection processing is performed with a high-level component, and then the detection completion can be realized early by shifting to detection processing with a low-level component. The amount of processing can be reduced. Furthermore, since the detection process is performed with a high-level component, the detection accuracy can be kept high.

  The present invention can be suitably used, for example, in the field of image processing for detecting a desired target object from an input image.

1 is a block diagram of an image processing apparatus according to Embodiment 1 of the present invention. The internal block diagram of the feature-value vector evaluation part in Embodiment 1 of this invention Processing flowchart of feature vector evaluation unit 5 in Embodiment 1 of the present invention Processing flowchart of feature vector evaluation unit in Embodiment 1 of the present invention (A) Image of face image in Embodiment 1 of the present invention (b) Image of non-face image in Embodiment 1 of the present invention Flow chart showing the flow of learning process Configuration diagram of learning unit for ensemble learning Flowchart of detection processing in Embodiment 1 of the present invention Processing diagram of detection processing in Embodiment 1 of the present invention 1 is a configuration diagram of an electronic device including an image processing device according to Embodiment 1 of the present invention. Block diagram of an image processing apparatus in Embodiment 2 of the present invention Internal block diagram of feature vector evaluation unit in embodiment 2 of the present invention Operation flow chart of feature vector evaluation unit in embodiment 2 of the present invention (A) Original image in Embodiment 2 of the present invention (b) Image image subjected to multi-resolution analysis in Embodiment 2 of the present invention Flow chart of reference parameter generation in Embodiment 2 of the present invention Block diagram of an image processing apparatus in Embodiment 2 of the present invention Block diagram of a conventional image processing apparatus Block diagram of a conventional image processing apparatus

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 Image processing apparatus 2 Input image 3 Partial image clipping part 4 Feature quantity vector calculation part 5 Feature quantity vector evaluation part 6 Image identification part 7 Correlation value calculation part 8 Similarity degree calculation part 9, 21 Evaluation value calculation part 10 Determination part 11 reference Database 12 Reference parameter 13 Target object reference vector 14 Non-target object reference vector 15 Weight coefficient 16 Detection result 20 Feature vector 22 Intermediate determination unit 23 Final determination unit 24 Reference parameter storage unit 25 Control parameter storage unit 101 CPU
102 ROM
103 Bus 104 RAM
105 HD
107 I / F
108 Cameras 601, 1301 Evaluation value calculation unit 602, 1302 Intermediate determination unit 603, 1303 Final determination unit 604, 1304 Reference parameter storage unit 605, 1305 Control parameter storage unit 1101 Reference database storage unit 1102 Filter unit 1103 Learning processing unit 1104 Evaluation data Part 1104 evaluation part 1703 feature vector evaluation part (left direction)
1704 Feature vector evaluation unit (front)
1705 Feature vector evaluation unit (right direction)

Claims (16)

An image processing apparatus for detecting a target object from an input image,
A reference object having a plurality of reference parameters that are combinations of a reference vector including a target object reference vector that is an image vector of the target object and a non-target object reference vector that is an image vector of a non-target object and a weight coefficient of the reference vector. A database,
A partial image cutout unit for cutting out a partial image from the input image;
A feature amount vector calculation unit that calculates a feature amount vector of the partial image cut out by the image cutout unit;
A feature quantity vector evaluation unit that performs a detection process of detecting a target object from the partial image using the feature quantity vector and the reference parameter;
The feature vector evaluation unit includes an image identification unit that determines inclusion / non-inclusion of the target object in the partial image based on a calculation result calculated using the feature vector and the reference parameter;
An image processing apparatus that ends the detection process when the image identification unit determines that the target object is not included. The image processing apparatus according to claim 1, wherein the detection process is terminated when there are a plurality of image identification units and at least one of the plurality of image identification units determines that the target object is not included. A correlation value calculation unit for calculating a correlation value between a target object correlation value between the feature quantity vector and the target object reference data, a correlation value between the feature quantity vector and the non-target object reference data, and the correlation value. An evaluation value calculation unit that calculates an evaluation value that is a product of the target object correlation value or the non-target object correlation value calculated by the calculation unit and the weighting factor; and the target object of the partial image according to the evaluation value The image processing apparatus according to claim 1, further comprising a determination unit that performs inclusion / non-inclusion determination. An intermediate determination unit in which the determination unit repeatedly performs inclusion / non-inclusion determination using the different reference parameters when the calculation processing in the evaluation value calculation unit is a predetermined number of times or less, and calculation in the evaluation calculation unit The image processing apparatus according to claim 3, further comprising a final determination unit that performs inclusion / non-inclusion determination when the processing is performed a predetermined number of times or more. The feature quantity vector evaluation unit continues detection processing using the different reference parameters when at least one of the plurality of image determination units determines that the target object is included. The image processing apparatus described. The image processing apparatus according to claim 1, wherein the feature vector evaluation unit uses a plurality of different reference parameters in descending order of weighting factors. The image processing apparatus according to claim 1, wherein the feature quantity vector evaluation unit uses the target object reference data in ascending order of the number of dimensions of the non-target object reference data regarding the non-target object. The image processing apparatus according to claim 1, wherein the reference parameter is generated by ensemble learning from the target object reference vector and the non-target object reference vector. The image processing apparatus according to claim 8, wherein the ensemble learning uses a boosting technique. 10. The image according to claim 1, wherein the feature vector is calculated by a filter process that calculates at least one of an edge feature and a multi-resolution feature (hereinafter referred to as “wavelet feature”) by Haar wavelet transform. Processing equipment. The image processing apparatus according to claim 10, wherein the wavelet feature is obtained by octave division. The image processing apparatus according to claim 11, wherein the feature vector evaluation unit sequentially uses high-level components obtained by the octave division. The image processing apparatus according to claim 1, wherein the input image is at least one of encoded and decoded in a wavelet-compatible format. The image processing apparatus according to claim 13, wherein the wavelet-compatible format corresponds to a JPEG2000 format. An image processing method for detecting a target object from an input image,
Parameter setting that sets a plurality of parameters that are combinations of a reference vector including a target object reference vector that is an image vector of the target object and a non-target object reference vector that is an image vector of a non-target object and a weighting factor of the reference vector Steps,
A partial image cutout step of cutting out a partial image from the input image;
A feature amount vector calculating step of calculating a feature amount vector of the clipped partial image;
Using the feature vector and the parameter, a feature vector evaluation step for performing a detection process for detecting a target object from the partial image,
The feature vector evaluation step includes an image determination step for determining inclusion / non-inclusion of the target object in the partial image based on a calculation result calculated using the feature vector and the parameter;
An image processing method for ending the detection process when the plurality of image determination steps determine that the target object is not included. A correlation value calculating step in which the image determining step calculates a target object correlation value between the feature quantity vector and the target object reference data or a correlation value between the feature quantity vector and the non-target object reference data; and the correlation value The image processing method according to claim 15, further comprising: an evaluation value calculating step for calculating an evaluation value that is a product of the weight coefficient and the weight coefficient.

Images