JP2020038731A - Image recognition device, image recognition method, and image recognition program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a technique for further improving the recognition accuracy of a face image including a shielded area without increasing the cost for constructing a system and the processing load at the time of identification. An image recognition device 1 includes a determination unit 11 and an identification unit 12. The determination unit 11 determines a shielded area that is shielded in the input face image, which is an image representing the face. The identification unit 12 excludes the exclusion region associated with the shielding pattern based on the distribution of the shielding region from the input face image, and identifies the input face image. [Selection diagram] Fig. 1

Description

  The present invention relates to a technique for identifying a face image including a shielding area.

  An image recognition device for identifying a face image is well known. In such an image recognition device, when a part of the face is blocked by an attached object such as glasses, sunglasses, or a mask in an input face image or a registered face image, recognition accuracy is reduced.

  One technique related to such a problem is described in Patent Document 1. The related art described in Patent Literature 1 calculates an inter-pattern distance value between a plurality of small regions in an input image and each small region of a registered image at a position corresponding to each small region. Then, this related technique identifies an input image based on an integrated distance value calculated using only a predetermined number in ascending order of the inter-pattern distance values. In this related technique, a small area having a large inter-pattern distance value between an input image and a registered image does not contribute to identification. Therefore, according to this related technique, for a small area in which an input image and a registered image are largely different due to occlusion, it is possible to perform matching while avoiding the small area.

  Another technique related to such a problem is described in Patent Document 2. In the related art described in Patent Literature 2, a plurality of pieces of face information having different wearing states of wearing objects (such as glasses) are registered in advance for each user. Then, this related technique specifies candidate face information similar to the input face image from among a plurality of pieces of face information. Furthermore, this related technology specifies different type face information having a different mounting state from the candidate face information among a plurality of face information registered for the user corresponding to the candidate face information. Then, in the related technique, when the input face image and the different-type face information are similar in an area other than the mounting area of the mounted object, the input face image is determined to represent the user corresponding to the candidate face information. As described above, according to this related technique, the collation can be performed regardless of the mounting state of the mounted object in the input face image.

  Another technique related to such a problem is described in Patent Document 3. In the related art described in Patent Literature 3, when it is determined that there is a wearing object in a face image, a first feature amount is extracted from a region avoiding the wearing object region. In this related technique, a face image with an attached object is registered together with the first feature amount. Further, in the related art, when it is determined that there is no wearing object in the face image, the second feature amount is extracted from the region including the assumed wearing region in addition to the first feature amount. In this related technique, a face image without a wearing object is registered together with the first feature amount and the second feature amount. This related technique extracts a first feature amount or a second feature amount from an input face image and performs collation with the first feature amount or the second feature amount of a registered image. As described above, this related technology also performs matching when there is a wearing object in the registered face image and no wearing object in the input face image, or when there is no wearing object in the registered face image and there is a wearing object in the input face image. be able to.

  Another related technique relating to such a problem is described in Patent Document 4. The related art described in Patent Literature 4 divides an input face image and a reference face image into a plurality of small areas each having a feature point as a vertex, and compares each of the corresponding small areas. Then, this related technique determines whether or not each small area of the input face image is a shielding area based on the comparison result.

  Another related technique related to such a problem is described in Patent Document 5. In the related art described in Patent Literature 5, a partially occluded area of an input face image is complemented by an associative memory circuit that has been learned in advance. Then, this related technique performs collation using the complemented recall image. As described above, in the related technique, all images are collated as images without occluded regions.

Patent No. 4803214 JP 2014-115784 A JP 2007-280250 A JP 2011-60038 A JP 2007-148873 A

  However, the above related art has the following problems.

  In the related art described in Patent Literature 1, even when the face of each of the input image and the registered image is different from each other, the arrangement and the contents of the shielding areas in each image may be similar. In such a case, this related technique uses an occluded region having a small inter-pattern distance value for matching. For this reason, this related technique may determine that the input image and the registered image indicating different persons actually indicate the same person.

  In the related art described in Patent Literature 2, it is necessary to register a plurality of face images having different wearing states for each user. However, not all users can register such a plurality of face images. Further, when the number of types of assumed wearing objects increases, the user must register face images having different wearing states for each of the wearing objects. This creates a problem in terms of convenience.

  Further, in the related art described in Patent Literature 3, when the types of assumed attachments increase, it is necessary to extract a first feature amount of each attachment that avoids the assumed attachment area. Therefore, in this related technique, it is necessary to extract the first feature amount of the pattern that differs by the number of types of assumed wearing objects, and the processing load increases. Note that if the first feature amount is extracted from a region avoiding the assumed attachment region of all assumed attachments, the number of patterns of the first feature amount does not increase even if the number of assumed attachment types increases. . However, in this case, as the number of types of assumed wearing objects increases, the area that can be used for matching out of the areas representing the face decreases. As a result, recognition accuracy decreases. In addition, as the number of types of attachments assumed increases, the cost of pre-learning the function of determining the presence / absence of attachments also increases.

  The related art described in Patent Literature 4 determines, for each small area of an input image, whether or not the small area of a reference face image is an occluded area based on a difference in luminance from a small area of a reference face image. However, due to the influence of illumination variation, the difference in luminance from the reference face image may be large even in a small area, even if the area is not a shielding area. In this case, this related technique determines that a small area that is not a shielding area is a shielding area. In addition, a small area of the input face image is covered by a shield having a small texture, and the texture in the corresponding small area of the reference face image may be originally small. At this time, the difference in luminance between such small areas may be small. In this case, the related art determines that the small area that is the shielding area is not the shielding area. For example, the difference in luminance between a part of the mask of a person wearing a white mask and the skin of the reference face may be small. In this case, the related art determines that the area shielded by the mask is not the shielded area. As described above, this related technique has a problem in the accuracy of determining a shielding area for each small area.

  Further, the related art described in Patent Document 5 cannot cope with a case where a registered image has a shielding area. This is because learning for generating a recollection image that complements a partial occlusion region requires a registered image without occlusion regions. Further, in the related art, when an input face image representing a person other than the person represented by the registered image is input, a recollection image close to the person represented by the registered image is generated. Therefore, this related technique may determine that such an input face image actually indicates the same person as a registered face image indicating a different person.

  The present invention has been made to solve the above problems. That is, an object of the present invention is to provide a technique for further improving the recognition accuracy of an image including an occluded region without increasing the cost for system construction and the processing load at the time of identification.

  An image recognition device according to the present invention compares an image representing a face (face image) with an image representing a standard face (standard face image) to determine a face occlusion pattern in the face image. And an identification unit that identifies the face image by excluding an area based on the shielding pattern in the face image.

  Also, the image recognition method of the present invention determines a face occlusion pattern in the face image by comparing an image representing a face (face image) with an image representing a standard face (standard face image). And identifying the face image by excluding an area based on the shielding pattern in the face image.

  Further, the image recognition program of the present invention determines a face occlusion pattern in the face image by comparing an image representing a face (face image) with an image representing a standard face (standard face image). The computer device is caused to execute a determination step and an identification step of identifying the face image by excluding an area based on the shielding pattern in the face image.

  The present invention can provide a technique for further improving the recognition accuracy of an image including an occluded region without increasing the cost of system construction and the processing load at the time of identification.

FIG. 1 is a functional block diagram of an image recognition device according to a first embodiment of the present invention. FIG. 1 is a diagram illustrating an example of a hardware configuration of an image recognition device according to a first embodiment of the present invention. 5 is a flowchart illustrating an operation of the image recognition device according to the first embodiment of the present invention. It is a functional block diagram of an image recognition device as a second embodiment of the present invention. FIG. 14 is a diagram schematically illustrating an example of a small area obtained by dividing a face image according to the second embodiment of the present invention. It is a figure which shows typically another example of the small area | region where the face image was divided in 2nd Embodiment of this invention. It is a figure which shows typically another example of the small area | region where the face image was divided in 2nd Embodiment of this invention. It is a figure which shows typically another example of the small area | region where the face image was divided in 2nd Embodiment of this invention. 9 is a flowchart illustrating an operation of the image recognition device according to the second embodiment of the present invention. It is a functional block diagram of an image recognition device as a third embodiment of the present invention. 13 is a flowchart illustrating an operation of the image recognition device according to the third embodiment of the present invention. It is a functional block diagram of an image recognition device as a fourth embodiment of the present invention. 13 is a flowchart illustrating an operation of the image recognition device according to the fourth embodiment of the present invention.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

(First Embodiment)
FIG. 1 shows a configuration of an image recognition device 1 according to a first embodiment of the present invention. In FIG. 1, the image recognition device 1 includes a determination unit 11 and an identification unit 12.

  Here, as shown in FIG. 2, the image recognition device 1 includes a CPU (Central Processing Unit) 1001, a RAM (Random Access Memory) 1002, a ROM (Read Only Memory) 1003, and a storage device 1004 such as a hard disk. , An imaging device 1005, and an output device 1006. The ROM 1003 and the storage device 1004 store a computer program and various data for causing a computer device to function as the image recognition device 1 of the present embodiment. The imaging device 1005 is a device such as a camera or a video camera that generates an image representing a face by imaging a person's face. The output device 1006 is a device such as a display that outputs information. The CPU 1001 controls each unit of the computer device by reading the computer program and various data stored in the ROM 1003 and the storage device 1004 into the RAM 1002 and executing them.

  In this case, the determination unit 11 includes an imaging device 1005 and a CPU 1001 that reads a computer program and various data stored in a ROM 1003 and a storage device 1004 into a RAM 1002 and executes the data. The identification unit 12 includes an output device 1006 and a CPU 1001 that reads a computer program and various data stored in a ROM 1003 and a storage device 1004 into a RAM 1002 and executes the data. Note that the hardware configuration of the image recognition device 1 and its functional blocks is not limited to the above-described configuration.

  Next, each functional block of the image recognition device 1 will be described.

  The determination unit 11 determines a face occlusion pattern in the face image by comparing the face image and the standard face image. Here, the face image is an image representing a face. For example, the determination unit 11 acquires a face image via the imaging device 1005. The standard face image is an image representing a standard face. The standard face image is determined in advance, and is stored, for example, in the storage device 1004. The face image and the standard face image are images obtained by cutting out a region representing a face from the corresponding image. In the present embodiment, it is assumed that the face image and the standard face image have substantially the same size, face direction, and the like.

  In addition, the shielding pattern is information indicating what area of the face has been blocked by what kind of blocking object. Specifically, the shielding pattern may be information such that the mouth is hidden by a mask or the eyes are hidden by sunglasses. For example, the determination unit 11 compares the face image and the standard face image to estimate a blocking area where the face is blocked by the blocking object, and determines a blocking pattern in consideration of the estimated blocking area distribution. You may.

  The identification unit 12 identifies a face image by excluding a region based on a shielding pattern in the face image. It is assumed that the exclusion area is determined in advance according to an assumed shielding pattern. Note that the exclusion area does not necessarily need to match the shielding area estimated in the process of determining the shielding pattern by the determination unit 21. In addition, as a face image identification technique, a known technique can be adopted. Then, the identification unit 12 outputs the identification result to the output device 1006.

  The operation of the image recognition device 1 configured as described above will be described with reference to FIG.

  First, the determination unit 11 compares an input face image (input face image) with a standard face image (step S1).

  Next, the determination unit 11 determines a shielding pattern in the input face image based on the comparison result of Step S1 (Step S2).

  Next, the identification unit 12 identifies the input face image by excluding an area based on the occlusion pattern determined in step S2 in the input face image. Then, the identification unit 12 outputs an identification result (Step S3).

  Thus, the operation of the image recognition device 1 ends.

  Next, effects of the first exemplary embodiment of the present invention will be described.

  The image recognition device according to the first embodiment of the present invention can further improve the recognition accuracy of an image including an occluded region without increasing the cost for constructing a system and the load of identification processing.

  The reason is that the determination unit determines the occlusion pattern of the face image by comparing the face image and the standard face image, and the identification unit excludes an area based on the occlusion pattern in the face image, This is because identification is performed.

  As described above, in the present embodiment, the face image is compared with the standard face image, and the occlusion pattern is determined in consideration of the distribution of occluded regions in the face image. In addition, the present embodiment does not exclude the occluded area itself that is estimated to be occluded by comparison with the standard face image, but defines the occluded pattern determined in consideration of the distribution of the occluded area. The discrimination is performed by excluding the region that has existed. As a result, the present embodiment realizes occlusion detection that is robust against an occlusion region determination error. Further, according to the present embodiment, it is possible to prevent the authentication accuracy from being lowered due to the occlusion region determination error, and to improve the face image recognition accuracy.

  Further, the present embodiment does not require a prior learning of a function of determining the presence or absence of a shielding object for each type of shielding object. In addition, the present embodiment does not require pre-registration of a plurality of images corresponding to the presence or absence of shielding for each user. Further, in the present embodiment, it is not necessary to extract feature amounts corresponding to the number of types of shields. In the present embodiment, a standard face image is prepared in advance, and an area to be excluded may be determined according to an assumed shielding pattern, which may increase the cost of system construction and the load of identification processing. Absent.

(Second embodiment)
Next, a second embodiment of the present invention will be described in detail with reference to the drawings. In each of the drawings referred to in the description of the present embodiment, steps having the same configuration and the same operation as those in the first embodiment of the present invention are denoted by the same reference numerals, and detailed description in the present embodiment will be given. Description is omitted.

  First, the configuration of an image recognition device 2 according to a second embodiment of the present invention is shown in FIG. In FIG. 4, the image recognition device 2 differs from the image recognition device 1 according to the first embodiment of the present invention in that a determination unit 21 replaces the determination unit 11, and an identification unit 22 replaces the identification unit 12. And a dividing unit 23 and a reliability calculating unit 24.

  Here, the image recognition device 2 can be configured by the same hardware elements as the image recognition device 1 according to the first embodiment of the present invention described with reference to FIG. In this case, the dividing unit 23 includes an imaging device 1005 and a CPU 1001 that reads a computer program and various data stored in a ROM 1003 and a storage device 1004 into a RAM 1002 and executes the data. Further, the reliability calculation unit 24 is configured by a CPU 1001 that reads a computer program and various data stored in the ROM 1003 and the storage device 1004 into the RAM 1002 and executes the data. Note that the hardware configuration of the image recognition device 2 and its functional blocks is not limited to the above-described configuration.

  Next, each functional block of the image recognition device 2 will be described.

  The dividing unit 23 divides the face image into small areas (also called patches). Note that the dividing unit 23 may acquire a face image via the imaging device 1005. For example, the dividing unit 23 may divide the face image into small regions having the same shape. Further, the dividing unit 23 may divide the face image such that the shape of a certain small area is different from at least another part. Further, each of the divided small areas may have a portion that overlaps with another small area. It is desirable that the dividing unit 23 divides the face image into small areas having the same arrangement and shape as the standard face image. For example, the dividing unit 23 may divide the standard face image into small areas in advance, and divide the input face image into similar small areas. Alternatively, the dividing unit 23 may acquire information of a standard face image divided in advance into small areas and divide the face image into small areas similar to the acquired small areas.

  Note that the dividing unit 23 may perform a process of normalizing the face image before dividing the face image. This is to deal with a case where the face image input to the image recognition device 2 is not normalized. Specifically, the dividing unit 23 generates a face image in which the face in the face image is normalized. Note that the normalization here refers to performing a deformation process so that the face angle, size, and the like become a predetermined angle and size. For example, the dividing unit 23 may detect a feature point such as an eye or a mouth in the input face image and perform geometric transformation so that the feature point is located at a predetermined position. When the image input to the image recognition device 2 is an image representing a wider range than the face, the dividing unit 23 may perform normalization after cutting out the face image from the corresponding image. For example, the dividing unit 23 may employ the method of Japanese Patent No. 4653606 for normalizing the face image. Note that the dividing unit 23 may normalize the input face image using another method.

  As a specific example, for example, the dividing unit 23 may divide the normalized face image into small lattice-like rectangular small areas as shown in FIG. Alternatively, as shown in FIG. 6, some or all of the small areas may be divided so as to overlap with other small areas. In FIG. 6, the small area indicated by the bold solid line rectangle overlaps the four small areas indicated by the dashed rectangle. Alternatively, the dividing unit 23 may divide the normalized face image such that the size of a certain small area is different from the size of at least some other small areas, as shown in FIG. Alternatively, the dividing unit 23 may divide the normalized face image into small regions having shapes other than rectangles (for example, triangles), as shown in FIG. The division unit 23 may divide the image into small areas of any shape and size as long as the processing of the reliability calculation unit 24 and the identification unit 22 is not hindered.

  The reliability calculation unit 24 compares the corresponding small areas of the small area forming the face image with those of the small area forming the standard face image. Then, the reliability calculation unit 24 calculates the reliability based on the comparison result for each small area of the face image. The reliability indicates the possibility that the small area is an occluded area where the face is hidden by an occluder. The reliability may be smaller as the possibility that the small region is the shielding region is higher, and may be larger as the possibility is lower. Conversely, the reliability may be larger as the possibility that the small region is a blocking region is higher, and may be smaller as the probability is lower.

  The standard face image is stored in the storage device 1004 in advance. It is assumed that the standard face image has been normalized in advance. For example, the standard face image may represent a face like a general ID photo that does not include the occluded area. In addition, it is desirable to adopt a face image that is easily collated with more face images as the standard face image. For example, an image representing an average face obtained by averaging the faces of a plurality of persons may be used as the standard face image. As the standard face image, an image representing Wolf / Ram (a sample whose similarity with the faces of a plurality of persons satisfies a predetermined condition) may be used. As the standard face image, an image representing an object (artificial object) artificially created to be similar to the faces of a plurality of persons may be used. Further, the standard face image may be divided into small areas in advance. In this case, the small area constituting the standard face image may be one divided by the dividing unit 23.

  For example, the reliability calculating unit 24 may calculate the reliability based on the sum of the differences between the luminance values between the corresponding small areas. However, the reliability using the luminance value is easily affected by illumination fluctuation. Therefore, the reliability calculation unit 24 may calculate the reliability using the feature amount vector for each small area. The feature amount vector may be, for example, information obtained by a feature extraction process such as a Gabor filter or LBP (Local Binary Pattern) often used in face authentication. Alternatively, the feature amount vector may be information obtained by a feature extraction process such as a Haar-like filter often used in face detection. Then, in this case, the reliability calculation unit 24 can calculate the reliability based on the distance or the correlation between the feature vectors between the corresponding small regions. Based on such a feature amount vector, the reliability calculation unit 24 can calculate the reliability in which the influence of illumination fluctuation or the like is reduced. In the case where the reliability based on the distance between the feature amount vectors is calculated, the larger the value of the reliability, the higher the possibility that the area is the occlusion area. Further, when the reliability based on the correlation between the feature amount vectors is calculated, the smaller the value of the reliability, the higher the possibility that the region is the occluded region. For example, the reliability calculation unit 24 may calculate a normalized correlation value between the feature amount vectors and use the calculated correlation value as the reliability of each small region. In this case, it can be considered that the closer to 1 the reliability is, the higher the possibility of the non-blocking region, and the closer to 0, the higher the possibility of the blocking region. Hereinafter, an example in which the reliability is a normalized correlation value will be mainly described.

  The determining unit 21 determines the occlusion pattern based on the distribution of the reliability of each small area in the face image. Here, the determination unit 21 may determine whether or not the distribution of the reliability of the small region group forming the middle region satisfies the shielding condition with respect to the middle region obtained by combining the small region groups in the predetermined range. . Then, when the shielding condition is satisfied, the determination unit 21 may estimate that the middle region is the shielding region. Then, the determining unit 21 may determine the shielding pattern based on the estimated distribution of the shielding area.

  For example, the determination unit 21 may determine whether the reliability of each small region constituting the middle region is higher or lower than a threshold. Then, the determination unit 21 may regard the reliability of the small region having a different determination result from the determination result of the surrounding small region as noise. Specifically, the determination unit 21 determines that the small region in which the reliability of the small region group forming the middle region is lower than the threshold is a majority of the small region group or more, May be estimated to be a shielding area. Note that the shielding condition may be another condition. Then, the determining unit 21 may determine the shielding pattern based on the distribution of the middle region estimated as the shielding region. Note that the middle region may be set in a stepwise manner. For example, the determination unit 21 may have a configuration in which a middle area obtained by combining small area groups in a predetermined range is further combined into a predetermined area to form a large area. In such a case, the determination unit 21 may estimate that the large area is a shielded area when the distribution of the medium areas constituting the large area that is estimated to be the shielded area satisfies the shielding condition. . Then, in this case, the determination unit 21 may determine the shielding pattern based on the distribution of the large area estimated to be the shielding area. Further, the present invention is not limited to the three stages of the small region, the middle region, and the large region, and the small region to the large region may be designed in more stages.

  For example, the determination unit 21 may set the upper face area in which the small area groups above the nose in the face image are put together as the middle area. In this case, when the distribution of the reliability of the small area group in the upper face area satisfies the shielding condition, the determination unit 21 may determine that the pattern is a shielding pattern using sunglasses. In addition, the determination unit 21 may set the lower face area in which the small area groups below the nose are put together as the middle area. In this case, when the distribution of the reliability of the small area group in the lower face area satisfies the shielding condition, the determination unit 21 may determine that the pattern is a masking mask pattern. In addition, the determination unit 21 may set the middle area according to various types of assumed shielding patterns.

  Here, the merit of determining the shielding pattern based on the distribution of the reliability of the small region or the distribution of the shielding region will be described. The result of the reliability calculated for each small area is likely to be unstable depending on the shooting conditions. For this reason, if the occlusion determination based on the reliability is performed for each small area, an error is more likely to occur. Note that the error here means that a small area that is originally a shielding area is determined to be not a shielding area, and a small area that is not originally a shielding area is determined to be a shielding area. Also, the reliability at which such an error can occur can be said to be noise. If it is assumed that a region other than the shielding region is used for identification based on such a shielding determination for each small region, there is a possibility that a small region (shielding region) that is not originally suitable for identification is used for identification as a non-shielding region. is there. In addition, there is a possibility that a small area (an area that is not a shielding area) that is originally suitable for an identification target is not used for identification as a shielding area. Therefore, the occlusion determination based on the reliability of each small area reduces the authentication accuracy.

  On the other hand, masks and sunglasses are typical examples of facial masking patterns that often occur in daily life. These shielding patterns are for shielding a very large area of about 1/3 or 1/2 of the face. Considering such characteristics of the shielding pattern, if it is assumed that there is no noise in the reliability, the small regions having the reliability lower (or higher) than the threshold are included in the small region group having the reliability higher (or lower) than the threshold. It is unlikely that the region is slightly present. Therefore, when a reliability that is clearly different from the reliability of the surrounding small area is calculated, the reliability can be considered to be noise.

  Therefore, as described above, the determination unit 21 considers such an error (noise) and determines whether or not the distribution of the reliability in the middle region obtained by combining the small region groups in the predetermined range satisfies the shielding condition. Then, it is estimated whether or not the middle area is a shielding area. Accordingly, the determination unit 21 can estimate the shielding area with higher accuracy than when determining whether or not the small area is a shielding area using only the reliability. As a result, the determination unit 21 can accurately determine the shielding pattern.

  The identification unit 22 identifies each small area in the input face image other than the exclusion area based on the shielding pattern. It is assumed that the exclusion area is determined in advance according to the shielding pattern, as in the first embodiment of the present invention. For example, in the case of a masking pattern using a mask, the lower half area of the face may be defined as an exclusion area. In addition, in the case of a shielding pattern using sunglasses, the upper half area of the face may be defined as the exclusion area. Note that the exclusion area does not necessarily need to match the shielding area estimated in the process of determining the shielding pattern by the determination unit 21.

  Then, the identification unit 22 identifies the face image based on the identification result for each small area other than the exclusion area. For example, the identification unit 22 may calculate an identification score for each small area other than the exclusion area based on the occlusion pattern in the face image. In that case, the identification unit 22 may calculate an integrated score obtained by integrating the calculated identification scores of the respective small areas, and output the integrated score as the identification result.

  Specifically, the identification unit 22 compares the small areas obtained by dividing the input face image with the small areas constituting the registered face image by comparing the corresponding small areas with each other. A score may be calculated. In this case, the identification unit 22 compares the corresponding small areas in an area that is not included in any of the exclusion areas based on each shielding pattern in the input face image and the registered image. Then, the identification unit 22 calculates an identification score for each small area other than the exclusion area based on the comparison result between the small areas. Then, the identification unit 22 integrates the identification scores, calculates an integrated score, and outputs the integrated score.

  Here, it is assumed that the registered face image is stored in the storage device 1004 in advance. Further, the storage device 1004 may store information indicating a shielding pattern of the registered face image together with the registered face image. Further, the storage device 1004 may store information indicating a small area constituting the registered face image together with the registered face image. The shielding pattern of the registered face image may be information determined by the determining unit 21 when the registered face image is registered. Further, the small area constituting the registered face image may be information divided by the dividing unit 23 when the registered face image is registered.

  In addition, the identification unit 22 may use a value based on a distance or a correlation between feature amount vectors as an identification score between corresponding small regions. In this case, the feature amount vector of each small region may be information obtained by a feature extraction process such as Gabor fill or LBP. However, it is preferable that the identification unit 22 uses a feature amount vector having a higher identification ability than the feature amount vector used in calculating the reliability with the standard face image in the identification process with the registered face image. For example, the identification unit 22 may learn a low-dimensional conversion matrix by performing a linear discriminant analysis (Linear Discriminant Analysis) on the feature amount vector extracted from the learning data. In this case, the learning data may be a face image correctly labeled for each person. Such learning data may be a registered face image. For example, in a case where all users of the image recognition apparatus 2 can be identified (for example, a case used for entrance / exit management), it is desirable to use registered face images of all system users as learning data. Alternatively, the learning data is not limited to the registered face image, and may be another labeled face image. For example, in a case where the image recognition device 2 is used for authentication of an unspecified number of persons (for example, a case where it is used for detecting a suspicious individual), as the learning data, regardless of whether the user is a system user or not, the labeled data is used. It is desirable to use face images of a large number of humans. Then, the identification unit 22 generates a low-dimensional feature amount vector with improved identification ability by applying the above-described conversion matrix to the feature amount vector for each small region of the input face image and the registered face image. Good.

  The operation of the image recognition device 2 configured as described above will be described with reference to FIG. It is assumed that a standard face image and a registered face image are stored in the storage device 1004 in advance. It is also assumed that the standard face image has been divided into small areas in advance. It is also assumed that the registered face image has been divided into small areas in advance. Also, it is assumed that the registered face image is stored together with information indicating the shielding pattern.

  9, first, the dividing unit 23 normalizes the input face image (input face image) (step S21).

  Next, the dividing unit 23 divides the input face image normalized in step S21 into small areas (step S22).

  Next, the reliability calculation unit 24 compares the corresponding small areas between each small area of the face image divided in step S22 and each small area of the standard face image. Thereby, the reliability calculating unit 24 calculates the reliability of each small area (Step S23).

  Next, the determining unit 21 determines the shielding pattern of the input face image based on whether or not the distribution of the reliability satisfies the shielding condition for the small region group forming the middle region (Step S24).

  Specifically, as described above, when the distribution of the reliability of the middle region representing the upper half or the lower half of the face satisfies the shielding condition, the determination unit 21 estimates that the middle region is the shielding region. On the other hand, if the distribution of the reliability of the middle region representing the upper half or the lower half of the face does not satisfy the shielding condition, the determination unit 21 estimates that the middle region is not the shielding region. Then, when it is estimated that such a middle region is a shielding region, the determination unit 21 may determine a shielding pattern corresponding to the set middle region. For example, when the determination unit 21 estimates that the upper half area is a shielding area, it determines that the pattern is a shielding pattern using sunglasses. In addition, when the determination unit 21 estimates that the lower half area is the shielding area, it determines that the lower half area is the shielding pattern by the mask.

  Next, the identification unit 22 collates the corresponding small areas between each of the small areas into which the input face image is divided and each of the small areas constituting the registered face image. Thereby, the identification unit 22 calculates an identification score between the small areas (Step S25).

  As described above, the identification unit 22 may calculate the identification score from the distance or the correlation between the feature vectors or the low-dimensional feature vectors.

  Next, the identification unit 22 calculates an integrated score obtained by integrating the identification scores between the small areas calculated in step S25, excluding the area based on the shielding pattern determined in step S24 (step S26).

  Here, when one or both of the corresponding small areas are included in the exclusion area based on the occlusion pattern of each image between the input face image and the registered face image, the identification unit 22 determines the identification score between the small areas. Not used for integration. Then, when both of the corresponding small regions are not included in the exclusion region based on the shielding pattern of each image, the identification unit 22 uses the identification score between the small regions for integration.

  For example, the identification unit 22 may use the average value of the identification scores of the corresponding small areas as the integrated score. If no occlusion pattern has been determined for both the input face image and the registered face image, the identification unit 22 sets the average value of the identification scores of all the small areas as the integrated score. For example, when a mask occlusion pattern is determined in one or both of the input face image and the registered face image, the identification unit 22 sets the lower half of the face as an exclusion area, and determines the identification score of the upper half small area. The average value may be used as the integrated score. Further, for example, the identification unit 22 may multiply the identification score of each target small region by the reliability of each small region, take an average, and use a weighted average value as an integrated score.

  As described above, the operation of the image recognition device 2 ends.

  Next, effects of the second exemplary embodiment of the present invention will be described.

  The image recognition device according to the second embodiment of the present invention can further improve the recognition accuracy of the face image including the occluded region without increasing the cost for constructing the system and the load of the identification process.

  The reason is that the dividing unit divides the face image into small regions, and the reliability calculating unit calculates the reliability of each small region by comparing the corresponding small regions between the face image and the standard face image. Because you do. Then, the determination unit estimates the occlusion area of the face image based on the distribution of the reliability of each small area, and determines the occlusion pattern based on the distribution of the occlusion area. Then, the identification unit excludes an area defined based on the occlusion pattern in the face image, identifies each small area, and identifies the entire face image based on each identification result.

  As described above, the present embodiment does not determine whether each of the small regions constituting the face image is a shielding region based on the reliability alone, but based on the distribution of the reliability over a wider range. Estimate the occluded area. As a result, in the present embodiment, it is possible to reduce the influence of an error in the occlusion determination for each small area, and to estimate a more accurate occluded area. In the present embodiment, the shielding pattern is determined based on the estimated distribution of the shielding region. Therefore, the influence of the error in the shielding determination can be further reduced, and the shielding pattern can be determined with higher accuracy. Then, in the present embodiment, the identification is performed by excluding the area estimated according to the determined shielding pattern, instead of excluding the area estimated as the shielding area as it is. For this reason, the present embodiment is more robust against occlusion determination errors. Then, in the present embodiment, such a region is excluded, and the face image is identified based on the identification result for each small region. As a result, the present embodiment can realize occlusion detection that is robust against occlusion region determination errors, and improve face image recognition accuracy.

  Furthermore, in the present embodiment, since the occlusion region and the occlusion pattern are determined by comparison with one standard face, a large amount of learning data required for the occlusion region and occlusion pattern determination by the existing technology is used. Do not need. As a result, this embodiment can significantly reduce the cost of system construction.

  In the present embodiment, an example has been described in which the registered image is stored in the storage device in advance together with the information on the shielding pattern and the divided small areas. However, the present embodiment is not limited to this, and the identification and the occlusion pattern of the registered image may be determined when the input face image is identified.

(Third embodiment)
Next, a third embodiment of the present invention will be described in detail with reference to the drawings. In each of the drawings referred to in the description of the present embodiment, steps having the same configuration and the same operation as those in the third embodiment of the present invention are denoted by the same reference numerals, and detailed description in the present embodiment will be given. Description is omitted.

  First, FIG. 10 shows the configuration of an image recognition device 3 according to a third embodiment of the present invention. In FIG. 10, the image recognition device 3 includes an overall identification unit 35 and a switching unit 36 in addition to the same configuration as the image recognition device 2 according to the second embodiment of the present invention.

  Here, the image recognition device 3 can be configured by the same hardware elements as the image recognition device 1 according to the first embodiment of the present invention described with reference to FIG. In this case, the overall identification unit 35 includes an output device 1006 and a CPU 1001 that reads a computer program and various data stored in a ROM 1003 and a storage device 1004 into a RAM 1002 and executes the same. The switching unit 36 is configured by a CPU 1001 that reads a computer program and various data stored in a ROM 1003 and a storage device 1004 into a RAM 1002 and executes the data. Note that the hardware configuration of the image recognition device 3 and its functional blocks is not limited to the above-described configuration.

  Next, each functional block of the image recognition device 3 will be described.

  The overall identification unit 35 identifies a face image using information representing the entire area of the face in the face image. Specifically, for example, the overall identification unit 35 obtains a feature vector by performing feature extraction from the entire face. Further, the overall identifying unit 35 also obtains a feature amount vector for the entire face from the registered face image. Then, the overall identification section 35 may calculate the overall identification score based on the feature vector of the entire face of the input face image and the feature vector of the entire face of the registered face image. In this case, the overall identification unit 35 outputs the overall identification score as an identification result.

  The switching unit 36 switches whether to use the identification unit 22 or the overall identification unit 35 to identify the input face image based on the determination result of the shielding pattern by the determination unit 21.

  The operation of the image recognition device 3 configured as described above will be described with reference to FIG. In the following, it is assumed that the registered face image is stored in the storage device 1004 together with the information indicating the shielding pattern and the small area.

  11, first, the image recognition device 3 operates in the same manner as the image recognition device 2 according to the second embodiment of the present invention in steps S21 to S24, and determines the occlusion pattern of the input face image.

  Next, the switching unit 36 switches subsequent processing based on the determined shielding pattern. Specifically, the switching unit 36 determines whether or not it is determined that at least one of the input face image and the registered face image has a shielding pattern regardless of the type (step S31).

  Here, when it is determined that at least one of the patterns has the shielding pattern, the switching unit 36 controls the identification unit 22 to execute steps S25 to S26 as in the second embodiment of the present invention. .

  On the other hand, if it is determined that neither the input face image nor the registered face image has the occlusion pattern, the switching unit 36 controls the overall identification unit 35 to execute the following step S32.

  Here, the overall identification unit 35 performs identification using the information on the entire face of the input face image and the information on the entire face of the registered face image. Specifically, the overall identification unit 35 calculates an overall identification score using the feature amount vector obtained from the entire face of the input face image and the feature amount vector obtained from the entire face of the registered face image (step). S32).

  For example, the overall identification unit 35 may obtain a feature vector from the entire face of each image by a feature extraction process such as a Gabor filter or LBP. Then, the overall identification unit 35 may calculate the overall identification score from the feature amount vectors of the entire face of each of the input face image and the registered face image and the distance or correlation therebetween.

  Thus, the operation of the image recognition device 3 ends.

  Next, effects of the third exemplary embodiment of the present invention will be described.

  The image recognition device according to the third embodiment of the present invention has the same effect as the second embodiment of the present invention, and further, does not reduce the authentication accuracy of the face image when there is no occluded area. Can be

  The reason is that if there is no occlusion pattern in either the input face image or the registered face image, the switching unit performs the identification using the information of the entire face without using the identification result for each small area. Is to switch. In addition, the switching unit switches the process to perform the identification using the identification result for each small area when at least one of the input face image or the registered face image has a shielding pattern.

  Here, when there is no occlusion pattern and it is not necessary to exclude a region, the classification result using information on the entire face tends to be more accurate than the integration of the classification result for each small region. Therefore, the present embodiment maintains the high authentication accuracy when at least one of the two images has the shielding pattern, and does not lower the matching accuracy even when there is no shielding pattern in both images.

  Note that, in the present embodiment, an example has been described in which the identification unit and the overall identification unit calculate the identification score based on the distance and the correlation between the feature amount vectors, but the identification unit employs another technology to perform the identification. May go.

(Fourth embodiment)
Next, a fourth embodiment of the present invention will be described in detail with reference to the drawings. In each of the drawings referred to in the description of the present embodiment, steps having the same configuration and the same operation as those in the third embodiment of the present invention are denoted by the same reference numerals, and detailed description in the present embodiment will be given. Description is omitted.

  First, FIG. 12 shows a configuration of an image recognition device 4 according to a fourth embodiment of the present invention. 12, the image recognition device 4 is different from the image recognition device 3 according to the third embodiment of the present invention in that an identification unit 42 is used instead of the identification unit 22 and a reliability level is used instead of the reliability calculation unit 24. The difference is that a calculation unit 44 and an overall identification unit 45 are provided in place of the overall identification unit 35, and a feature extraction unit 47 is further provided.

  Here, the image recognition device 4 can be configured by the same hardware elements as the image recognition device 1 according to the first embodiment of the present invention described with reference to FIG. In this case, the feature extraction unit 47 is configured by a CPU 1001 that reads a computer program and various data stored in a ROM 1003 and a storage device 1004 into a RAM 1002 and executes the same. The hardware configuration of the image recognition device 4 and its functional blocks is not limited to the above-described configuration.

  The feature extracting unit 47 extracts a feature vector for each small area of the face image divided by the dividing unit 23. The above-described various filters and techniques may be used to extract the feature amount vector. For example, the feature extracting unit 47 may extract a Gabor feature using a Gabor filter. More specifically, the feature extracting unit 47 calculates a feature amount vector for each of the small areas normalized and divided by the dividing unit 23 with respect to the input face image, and stores it in the storage device 1004. In addition, the feature extracting unit 47 may calculate a feature amount vector in advance for each small area for the standard face image and the registered face image. In the present embodiment, the standard face image only needs to be stored in the storage device 1004 as a feature amount vector for each small area, and the data of the face image itself need not be stored. Similarly, the registered face image only needs to be stored in the storage device 1004 as a feature amount vector for each small area, and the data of the registered face image itself need not be stored.

  The reliability calculation unit 44 is configured in substantially the same manner as the reliability calculation unit 24 according to the second and third embodiments of the present invention. That is, the reliability calculation unit 44 calculates the reliability between the input face image and the standard face image for each corresponding small area using the feature amount vector. However, the reliability calculating unit 44 is different from the reliability calculating unit 24 in that the feature amount vector of each small area for the input face image and the standard face image is obtained from the storage device 1004 instead of calculating.

  The identification unit 42 has substantially the same configuration as the identification unit 22 in the second and third embodiments of the present invention. That is, the identification unit 42 calculates an identification score between the input face image and the registered face image using the feature amount vector for each corresponding small area. However, the identification unit 42 is different from the identification unit 22 in that a feature amount vector of each small area is obtained from the storage device 1004 for the input face image and the registered face image instead of calculating the feature amount vector. Further, the identification unit 42 may previously learn a conversion matrix for converting a feature amount vector into a low dimension, as in the second and third embodiments of the present invention. In this case, the identification unit 42 calculates an identification score using a low-dimensional feature vector for each small area. For example, the identification unit 42 may calculate the normalized correlation value between the low-dimensional feature vectors of the corresponding small areas for the input face image and the registered face image as the identification score. Then, the identification unit 42 calculates an integrated score by excluding any areas based on the occlusion patterns of the input face image and the registered face image, and integrating the identification results of the small areas.

  The overall identification unit 45 calculates an overall identification score based on the connected feature vector of the input face image and the connected feature vector of the registered face image. Here, the connected feature vector is generated by connecting the feature vectors of the respective small areas. For example, it is assumed that feature amount vectors of n1 row and 1 column, n2 row and 1 column, and n3 row and 1 column have been extracted for three small regions. n1, n2, and n3 are positive integers, respectively. These may all be the same value, or at least some may be different from the others. In this case, the overall identifying unit 45 can generate a connected feature vector of (n1 + n2 + n3) rows and 1 column by connecting the feature vectors. That is, the number of dimensions of the connected feature vector is larger than that of the feature vector of each small area. The order of connection is arbitrary.

  Also, for the connected feature vector, the overall identification unit 45 also learns in advance a conversion matrix for converting the connected feature vector into a low-dimensional connected feature vector by linear discriminant analysis on the learning data. Is also good. In this case, the overall identification unit 45 may calculate the normalized correlation value between the low-dimensional connected feature vectors as the overall identification score.

  The operation of the image recognition device 4 configured as described above will be described with reference to FIG. In the following, it is assumed that the feature amount vector for each small area of the standard face image is stored in the storage device 1004. It is also assumed that the feature amount vector for each small area of the registered face image is stored in the storage device 1004 together with the shielding pattern determined by the determination unit 21 in advance.

  In FIG. 13, first, the dividing unit 23 executes steps S21 to S22 in the same manner as in the second embodiment of the present invention, and normalizes and divides the input face image.

  Next, the feature extracting unit 47 extracts a feature amount for each small area of the input face image and stores the feature amount in the storage device 1004 (step S41).

  Next, the reliability calculation unit 44 calculates the feature amount vector of each small area of the input face image stored in the storage device 1004 in step S41 and the feature amount vector of each small area of the standard face image stored in the storage device 1004. With the quantity vector, the reliability between the corresponding small areas is calculated (step S42).

  Next, the determining unit 21 executes Step S24 as in the second embodiment of the present invention, and determines a shielding pattern.

  Next, when the switching unit 36 determines that one or both of the input face image and the registered face image has the occlusion pattern, the identification unit 42 executes steps S43 and S26.

  Here, the identification unit 42 calculates the identification score between the corresponding small areas for the input face image and the registered face image using the feature amount vector of the small area stored in the storage device 1004 (step S43). .

  For example, as described above, the identification unit 42 may calculate the normalized correlation value as the identification score between the low-dimensional feature vectors obtained by converting the feature vectors of the corresponding small regions into low dimensions. Then, the identification unit 42 executes Step S26 as in the second embodiment of the present invention, and calculates an integrated score.

  On the other hand, when the switching unit 36 determines that neither the input face image nor the registered face image has the occlusion pattern, the overall identifying unit 45 executes steps S44 to S45.

  Here, the overall identifying unit 45 generates, for the input face image and the registered face image, a connected feature vector obtained by connecting the feature vectors of the small areas stored in the storage device 1004 (step S44).

  Then, the overall identification unit 45 calculates an overall identification score for the input face image and the registered face image by comparing the connected feature vectors (step S45).

  For example, as described above, the overall identification unit 45 may calculate the normalized correlation value between the low-dimensional connected feature amount vectors as the overall identification score after converting the connected feature amount vector into a low-dimensional one.

  Thus, the operation of the image recognition device 4 ends.

  Next, effects of the fourth embodiment of the present invention will be described.

  The image recognition device according to the fourth embodiment of the present invention realizes further efficient processing and higher speed while maintaining high authentication accuracy for a face image including an occluded area.

  The reason is that the feature extraction unit extracts and stores a feature amount vector for each of the small areas constituting the input face image, the standard face image, and the registered face image. Thus, the reliability calculation unit, the identification unit, and the overall identification unit can commonly use the already extracted feature amount vectors, and can increase the efficiency and speed of each process.

  In each of the second to fourth embodiments of the present invention described above, an example in which the identification unit and the overall identification unit identify an input face image by collating with a registered image has been mainly described. However, the identification unit and the overall identification unit in each embodiment may perform gender estimation, posture estimation, facial expression recognition, and the like of the person represented by the input face image without using the registered image. As described above, each of the embodiments can be applied to the use of performing various kinds of identification processing without using a registered image on a face image including an occluded region.

  In each of the above-described second to fourth embodiments of the present invention, the determination unit, the identification unit, and the overall identification unit determine the occlusion pattern using the distance or the correlation between the feature amount vectors and determine the input face. An example in which an image is identified has been mainly described. However, the present invention is not limited to this, and these functional blocks of the embodiments may calculate the reliability or the identification score by comparing the regions by another method. In this case, the feature extraction unit according to the fourth embodiment of the present invention may calculate and store information used for comparison for a small area of the corresponding image.

  Further, in each of the embodiments of the present invention described above, the shielding by the mask or the sunglasses is described as an example of the shielding pattern, but the type of the shielding pattern is not limited.

  In each of the above-described embodiments of the present invention, the input face image, the standard face image, and the registered face image indicate a part of a region representing a face included in the corresponding image. These face images may be images in which a face area has been cut out in advance, or may have been cut out during processing.

  Further, in each of the embodiments of the present invention described above, an example has been described centering on an example in which each functional block of the image recognition device is realized by a CPU that executes a computer program stored in a storage device or a ROM. The present invention is not limited to this, and some, all, or a combination of the functional blocks may be realized by dedicated hardware.

  In each of the embodiments of the present invention described above, the functional blocks of the image recognition device may be implemented by being distributed to a plurality of devices.

  In each of the embodiments of the present invention described above, the operation of the image recognition device described with reference to each flowchart may be stored in a storage device (storage medium) of a computer as an image recognition program of the present invention. Good. Then, the CPU may read and execute the computer program. In such a case, the present invention is constituted by the code of the computer program or the storage medium.

  In addition, the above-described embodiments can be implemented in appropriate combinations.

  Further, the present invention is not limited to the above-described embodiments, and can be implemented in various modes.

In addition, some or all of the above-described embodiments may be described as in the following supplementary notes, but are not limited thereto.
(Appendix 1)
A determination unit that determines a face occlusion pattern in the face image by comparing an image representing a face (face image) with an image representing a standard face (standard face image);
Excluding an area based on the shielding pattern in the face image, an identification unit that identifies the face image,
An image recognition device comprising:
(Appendix 2)
The determination unit determines the shielding pattern based on a distribution of reliability based on a comparison between each small area forming the face image and a small area corresponding to each small area of the face image in the standard face image. The image recognition device according to claim 1, wherein the determination is performed.
(Appendix 3)
When the distribution of the reliability of the small region group constituting the middle region satisfies a predetermined shielding condition for the middle region obtained by combining the small region groups in a predetermined range, the determination unit is configured based on the middle region. 3. The image recognition apparatus according to claim 2, wherein a shielding pattern is determined.
(Appendix 4)
The determination unit is configured to use the sunglasses when the distribution of the reliability satisfies the shielding condition in the middle region, where the upper face region in which the small region groups above the nose are grouped in the face image is the middle region. 4. The image recognition device according to claim 3, wherein the image recognition device determines that the pattern is a shielding pattern.
(Appendix 5)
The determination unit includes a mask lower face area in which the small area group below the nose is grouped in the face image as the middle area, and the distribution of the reliability in the middle area satisfies the shielding condition. The image recognition device according to Supplementary Note 3 or 4, wherein the image recognition apparatus determines that the pattern is a shielding pattern.
(Appendix 6)
A dividing unit that divides the face image into the small areas;
For the face image and the standard face image, a reliability calculation unit that calculates the reliability by comparing the corresponding small areas,
The image recognition device according to any one of Supplementary Notes 2 to 5, further comprising:
(Appendix 7)
The identification unit performs identification for each of the small areas other than the area based on the shielding pattern in the face image, and identifies the face image based on an identification result for each of the small areas. The image recognition device according to any one of supplementary notes 2 to 6.
(Appendix 8)
For the face image, an overall identification unit that performs identification using information on the entire face,
A switching unit that switches which of the identification unit and the overall identification unit is used based on a determination result of the determination unit,
The image recognition device according to any one of supplementary notes 1 to 7, further comprising:
(Appendix 9)
The image recognition device according to any one of supplementary notes 1 to 8, wherein a face image representing an average face obtained by averaging a plurality of human faces is used as the standard face image.
(Appendix 10)
The image recognition device according to any one of supplementary notes 1 to 8, wherein a face image whose similarity with a plurality of face images of a plurality of persons satisfies a predetermined condition is used as the standard face image.
(Appendix 11)
The image recognition device according to any one of supplementary notes 1 to 8, wherein a face image representing a standard facial artifact is used as the standard face image.
(Appendix 12)
The image recognition device according to any one of Supplementary Notes 1 to 11, wherein information represented by a feature vector is used as the face image and the standard face image.
(Appendix 13)
By comparing an image representing a face (face image) with an image representing a standard face (standard face image), a face occlusion pattern in the face image is determined,
An image recognition method for identifying the face image by excluding an area based on the shielding pattern in the face image.
(Appendix 14)
A determination step of determining a face occlusion pattern in the face image by comparing an image representing a face (face image) with an image representing a standard face (standard face image);
Excluding an area based on the shielding pattern in the face image, an identification step of identifying the face image,
Recognition program for causing a computer device to execute.

1, 2, 3, 4 Image recognition device 11, 21 Judgment unit 12, 22, 42 Identification unit 23 Division unit 24, 44 Reliability calculation unit 35, 45 Overall identification unit 36 Switching unit 47 Feature extraction unit 1001 CPU
1002 RAM
1003 ROM
1004 storage device 1005 imaging device 1006 output device

Claims (10)

A determining unit that determines a blocked area that is blocked in an input face image that is an image representing a face;
An exclusion region associated with a shielding pattern based on the distribution of the shielding region is excluded from the input face image, and an identification unit that identifies the input face image,
An image recognition device comprising:   The determination unit is configured to determine a reliability based on a comparison between a plurality of small regions constituting the input face image and a small region corresponding to each small region of the input face image in a standard face image representing a standard face. The image recognition device according to claim 1, wherein the occlusion pattern is determined based on a degree distribution.   The determination unit is configured such that, for a middle area which is a small area group obtained by combining the small areas in a predetermined range, when the distribution of the reliability of the small areas constituting the middle area satisfies a predetermined shielding condition, the shielding pattern The image recognition device according to claim 2, wherein the determination is made.   The determination unit sets the upper face area in which the small areas above the nose in the input face image are put together as the middle area. The image recognition device according to claim 3, wherein the image recognition device determines that the pattern is a shielding pattern.   The determining unit sets the lower face area in which the small areas below the nose in the input face image are grouped as the middle area, and, when the distribution of the reliability satisfies the shielding condition in the middle area, a mask. The image recognition device according to claim 3, wherein the image recognition device determines that the pattern is a shielding pattern. A dividing unit that divides the input face image into the small areas;
For the input face image and the standard face image, a reliability calculation unit that calculates the reliability by comparing the corresponding small areas,
The image recognition device according to any one of claims 2 to 5, further comprising:   The said identification part performs identification for every said small area other than the area | region determined based on the said shielding pattern in the said input face image, and performs identification of the said input face image based on the identification result for every said small area. The image recognition device according to any one of claims 2 to 6. For the input face image, an overall identification unit that performs identification using information on the entire face,
A switching unit that switches which of the identification unit and the overall identification unit is used based on a determination result of the determination unit,
The image recognition device according to any one of claims 1 to 7, further comprising: Determine an occluded area that is occluded in the input face image that is an image representing the face,
An image recognition method for identifying the input face image by excluding an exclusion area associated with a shielding pattern based on the distribution of the shielding area from the input face image. A process of determining an occluded area that is occluded in an input face image that is an image representing a face;
A process of excluding an exclusion region associated with a shielding pattern based on the distribution of the shielding region from the input face image, and identifying the input face image;
Recognition program for causing a computer device to execute.

Images