JP4561380B2 - Detection apparatus, detection method, and detection program - Google Patents

Description

  The present invention relates to a detection technique for detecting an object present in a target image.

  In recent years, many techniques for detecting a subject (for example, a human body or an object) existing in a target image have been proposed. There are simple methods such as skin color detection or edge detection, but when advanced processing such as authentication or state detection is executed after subject detection, a detection method with high detection position accuracy by template matching or the like is effective. .

  In template matching, there is a technique using a multi-resolution image in order to shorten the detection processing time (see Patent Document 1). This is a method of coarsely detecting and limiting the position using a low-resolution image obtained by reducing the detection target image at a predetermined reduction ratio, and performing more accurate detection with a high-resolution image based on the limited position information. Realizes high-speed and high-definition search processing.

JP 2000-134638 A

  However, in this method, only a subject having the same size as the template used for detection can be detected. Therefore, in order to detect a subject displayed in a plurality of sizes in the target image, a process for separately generating a multi-resolution image for detecting a plurality of sizes for each size of the detection target Time is required, and there is a problem that the speed cannot be sufficiently increased.

  Therefore, the present invention has been made in view of the above problems, and an object of the present invention is to provide a technique that can efficiently detect a plurality of sizes of subjects present in a target image.

  In order to solve the above problems, the invention of claim 1 is a detection device for detecting an object present in a target image, and is a first device used for detecting an object having a plurality of sizes in the target image. A first image group having a plurality of images with different magnifications relative to the target image, and a second image group used for detecting each of the plurality of objects in the target image at high speed. And a second image group that includes a plurality of images corresponding to a state in which the plurality of images in the first image group are reduced at a predetermined magnification, respectively. Image group creation means and object detection means for detecting the object using the hierarchical image group, wherein the first image group and the second image group are for a part of each image group For the target image As the magnification is identical to each other, and a part of the image the is created in shared state.

  According to a second aspect of the present invention, in the detection apparatus according to the first aspect of the invention, the hierarchical image group creating unit creates a plurality of reference images having different magnifications with respect to the target image, and the plurality of reference images. By creating a plurality of reduced images obtained by reducing each of them by a predetermined magnification, the hierarchical image group including the plurality of reference images and the plurality of reduced images is created, and a part of the plurality of reduced images includes: It is characterized by being created in common as images belonging to both the first image group and the second image group.

According to a third aspect of the present invention, in the detection apparatus according to the second aspect of the invention, the hierarchical image group creating means sets the first reduction rate α and the second reduction rate β to β = α ^N (N And a determination means for determining so as to satisfy a relationship of 2 or more) and a first reference image based on the target image and the first reference image between the images with the first reduction ratio α The plurality of images including the first reference image and the (N-1) intermediate layer reference images by sequentially reducing the reduction ratio to generate (N-1) intermediate layer reference images. And a means for creating a plurality of reduced images by further reducing each of the plurality of reference images at the second reduction ratio β.

  According to a fourth aspect of the present invention, in the detection apparatus according to the third aspect of the invention, the hierarchical image group creating means further includes one of the plurality of reduced images or each of the plurality of images. And means for creating at least one re-reduced image by reducing at a reduction ratio β of 2.

  The invention according to claim 5 is the detection apparatus according to claim 3 or claim 4, wherein the second reduction rate β is 1 / K (K is an integer of 2 or more). To do.

  According to a sixth aspect of the present invention, in the detection device according to any one of the first to fifth aspects, the object detection unit includes a target image selected from the first image group and a template. By using the comparison processing used, the plurality of sizes of objects are detected by repeating the operation of detecting the size of the target object in the target image, and the size of the object corresponding to the target image is detected. In operation, an object is detected by using the image of interest and an image obtained by reducing the image of interest at a predetermined magnification and sequentially using a low-resolution image from one image of the second image group. An object having a size to be detected in the attention image is efficiently detected.

  The invention of claim 7 is a detection method for detecting an object present in a target image, and is a first image group used for detecting an object having a plurality of sizes in the target image. A first image group having a plurality of images with different magnifications relative to the target image, and a second image group used for detecting each of the plurality of objects in the target image at high speed. A hierarchical image group creating step for creating a hierarchical image group including a second image group having a plurality of images corresponding to a state in which the plurality of images in the first image group are respectively reduced at a predetermined magnification; An object detection step of detecting the object using the hierarchical image group, wherein the first image group and the second image group are a magnification with respect to the target image with respect to a partial image of each image group Are identical to each other So that the characterized in that the part of the image the is created in shared state.

  The invention according to claim 8 is the detection method according to claim 7, wherein the hierarchical image group creating step creates a plurality of reference images having different magnifications with respect to the target image and the plurality of reference images. By creating a plurality of reduced images obtained by reducing each of them by a predetermined magnification, the hierarchical image group including the plurality of reference images and the plurality of reduced images is created, and a part of the plurality of reduced images includes: It is characterized by being created in common as images belonging to both the first image group and the second image group.

The invention according to claim 9 is the detection method according to claim 8, wherein the hierarchical image group creating step sets the first reduction rate α and the second reduction rate β to β = α ^N (N Is determined to satisfy the relationship of an integer of 2 or more, and a first reference image is created based on the target image, and the first reference image is between the images with the first reduction ratio α. The plurality of images including the first reference image and the (N-1) intermediate layer reference images by sequentially reducing the reduction ratio to generate (N-1) intermediate layer reference images. And a step of creating a plurality of reduced images by further reducing each of the plurality of reference images at the second reduction ratio β.

  Further, the invention of claim 10 is the detection method according to the invention of claim 9, wherein in the hierarchical image group creating step, one of the plurality of reduced images or each of the plurality of images is further processed. The method further includes the step of creating at least one re-reduced image by reducing at a reduction ratio β of 2.

  The invention according to claim 11 is the detection method according to claim 9 or claim 10, wherein the second reduction rate β is 1 / K (K is an integer of 2 or more). To do.

  According to a twelfth aspect of the present invention, in the detection method according to any one of the seventh to eleventh aspects, the object detection step includes a target image selected from the first image group and a template. By using the comparison processing used, the plurality of sizes of objects are detected by repeating the operation of detecting the size of the target object in the target image, and the size of the object corresponding to the target image is detected. In operation, an object is detected by using the image of interest and an image obtained by reducing the image of interest at a predetermined magnification and sequentially using a low-resolution image from one image of the second image group. An object having a size to be detected in the attention image is efficiently detected.

  The invention according to claim 13 is a first image group used for detecting an object having a plurality of sizes in the target image, and has a plurality of images having different magnifications with respect to the target image. A first image group, and a second image group used for detecting each of the plurality of size objects in the target image at high speed, wherein the plurality of images in the first image group are respectively A hierarchical image group creating step for creating a hierarchical image group including a second image group having a plurality of images corresponding to a state reduced at a predetermined magnification, and object detection for detecting the object using the hierarchical image group A program for detecting an object present in the target image by executing the step, wherein the first image group and the second image group are for a part of each image group. As the magnification for serial target image becomes the same as each other, and a part of the image the is created in shared state.

  The invention according to claim 14 is the program according to claim 13, wherein the hierarchical image group creating step creates a plurality of reference images having different magnifications with respect to the target image and each of the plurality of reference images. Are further reduced at a predetermined magnification to create the hierarchical image group including the plurality of reference images and the plurality of reduced images, and a part of the plurality of reduced images is It is characterized in that it is created in common as images belonging to both the first image group and the second image group.

Further, the invention according to claim 15 is the program according to claim 14, wherein the hierarchical image group creating step sets the first reduction rate α and the second reduction rate β to the computer, and β = α ^N (N is an integer equal to or greater than 2), a determination step of determining so as to satisfy the relationship, and a first reference image is created based on the target image and the first reference image is converted into the first reduction ratio α. The (N-1) intermediate layer reference images are sequentially reduced at the image reduction ratios to create the first reference image and the (N-1) intermediate layer reference images. Including a step of creating the plurality of reference images, and a step of creating a plurality of reduced images by further reducing each of the plurality of reference images at the second reduction rate β. .

  The invention of claim 16 is the program according to the invention of claim 15, wherein in the hierarchical image group creating step, one image or each of a plurality of images among the plurality of reduced images is stored in the computer. Further, the step of generating at least one re-reduced image is further executed by reducing at the second reduction rate β.

  The invention according to claim 17 is the program according to claim 15 or claim 16, wherein the second reduction ratio β is 1 / K (K is an integer of 2 or more). .

  According to an eighteenth aspect of the present invention, in the program according to any one of the thirteenth to seventeenth aspects, the object detection step causes the computer to select an image of interest selected from the first image group. The object having the size to be detected in the target image is detected by repeating the operation of detecting the object having the size corresponding to the target image by the comparison process using the template. In the operation of detecting the object, the object is detected by using the image of interest and the image of interest reduced at the predetermined magnification in order from the low resolution image of the second image group. By performing this, it is possible to efficiently detect an object having a size to be detected in the target image.

  According to the invention described in claims 1 to 18, an image group used for detecting an object of a plurality of sizes in the target image, an image group used for detecting an object in the target image at high speed, Since a hierarchical image group in which a part of the images is shared is created, an object having a plurality of sizes existing in the target image can be efficiently detected.

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

<Overview>
FIG. 1 is a diagram showing an outline of a detection apparatus (image processing apparatus) 1 according to an embodiment of the present invention. As shown in FIG. 1, the detection apparatus 1 is configured as a computer system (hereinafter also simply referred to as “computer”). Specifically, the computer (detection device) includes a CPU 2, a storage unit 3, a media drive 4, a display unit 5 such as a liquid crystal display, an input unit 6 such as a keyboard 6a and a mouse 6b that is a pointing device. And a communication unit 7 such as a network card. The storage unit 3 includes a plurality of storage media, specifically, a hard disk drive (HDD) 3a, and a RAM (semiconductor memory) 3b capable of processing at higher speed than the HDD 3a. The media drive 4 can read information recorded therein from a portable recording medium 9 such as a CD-ROM, DVD (Digital Versatile Disk), flexible disk, or memory card. The information supplied to the computer is not limited to the case where the information is supplied via the recording medium 9, but may be supplied via a network such as a LAN and the Internet.

  The detection apparatus 1 reads a software program (hereinafter simply referred to as “program”) read from the recording medium 9 and recorded in advance in the storage unit 3 onto a memory (RAM or the like), and loads the program on the CPU 2 or the like. By using and executing, various functions as described later are realized.

  FIG. 2 is a block diagram illustrating various functions included in the detection device 1. As shown in FIG. 2, the detection apparatus 1 includes an information setting unit 11, a pyramid image creation unit 12, a search window determination unit 13, and a matching processing unit 14.

  The information setting unit 11 has a function of setting reference information used for detection to information suitable for detection.

  The pyramid image creation unit 12 is an image group (hereinafter referred to as “multiple size detection use”) for detecting an object having a plurality of sizes present in a target image (hereinafter also referred to as “original image”) to be detected. And a group of images (hereinafter also referred to as “high-speed detection pyramid images”) that can detect a plurality of objects of a plurality of sizes existing in the target image at high speed. A hierarchical image group (hereinafter also referred to as “integrated pyramid image”). In the following description, the “multiple-size detection pyramid image”, the “high-speed detection pyramid image”, and the “integrated pyramid image” are also simply referred to as “pyramid images”, respectively.

  The search window determination unit 13 has a function of determining a search window region for raster scanning information used for detection on a predetermined image in the hierarchical image group created by the pyramid image creation unit 12 described above.

  The matching processing unit 14 has a function of performing a collation and search process in the search window region determined by the search window determination unit 13 and detecting a detection target object. Specifically, the matching processing unit 14 detects a target object existing in the target image by performing a comparison process using a template. The comparison process using the template includes (1) a process of comparing an image used as a template with the target image in pixel units (also referred to as a pattern matching process or a direct comparison process), and (2) a template image. There is a process (also referred to as an indirect comparison process) for comparing feature amounts extracted from the target image and the target image.

  FIG. 3 is a diagram showing an image PI showing several persons as subjects. FIG. 4 is a diagram showing an operation screen during program execution.

  The detection apparatus 1 having the above-described functions detects a plurality of sizes of objects present in the target image at high speed. For example, when detecting the face of a person existing in the image PI shown in FIG. 3 as a detection target, the detection apparatus 1 is not only a single size face but also in the operation screen Q shown in FIG. A plurality of sizes of faces within a range designated by the user (user) in the window W1 (from the smallest face to the largest face) can be detected.

  Here, an outline of the detection process executed in the detection apparatus 1 will be described.

  FIG. 5 is a diagram schematically illustrating a detection process executed in the detection apparatus 1. FIG. 6 is a schematic diagram showing the principle of detection processing for detecting an object having a plurality of sizes in the original image FE1 (PI) using a pyramid image. FIG. 7 is a schematic diagram showing the principle of detection processing for detecting a predetermined object existing in an image at high speed. FIG. 8 is a schematic diagram illustrating an example of a detection process in which both the detection processes illustrated in FIGS. 6 and 7 are combined, and is a diagram illustrating a comparative example of the detection process illustrated in FIG. The relationship between FIG. 5 and FIG. 8 will be described in detail later.

  As illustrated in FIG. 5, the detection apparatus 1 performs object detection by template matching using a template that is a detection technique with high detection position accuracy. Below, the outline | summary of each function which the detection apparatus 1 has is demonstrated.

  In the information setting unit 11, the detection apparatus 1 includes a detection template TE1, a reduction template TE2 obtained by reducing the detection template TE1 at a reduction rate β, and a re-reduction template TE3 obtained by further reducing the reduction template TE2 at a reduction rate β. The detection information IF (IF1) is created. Further, the pyramid image creation unit 12 creates a hierarchical image group MP (MP1) from the target image FE1 using the reduction rate γ, the reduction rate α, and the reduction rate β. Each reduction rate α, β, γ will be described later. Further, the search window determination unit 13 determines a search window region in an image of a predetermined hierarchy in the hierarchy image group MP1 created by the pyramid image creation unit 12 described above. This search window region is used when searching for a detection target object by a raster scan operation. Then, the matching processing unit 14 performs collation and search in the determined search window region, and performs detection processing of the detection target object. Specifically, as illustrated in FIG. 5, when detecting an object having a size corresponding to the detection template TE1 in the image FE2 in the second hierarchy, the detection apparatus 1 includes the re-reduction template TE3 and the sixth hierarchy. The above-described detection process is executed with the image FE6. If detected by the detection process, the detection process is performed between higher resolutions, that is, between the reduced template TE2 and the fourth-layer image FE4 based on the detection position. If further detection is performed, detection processing is performed between the detection template TE1 and the second-level image FE2 based on the detection position. According to such an operation, by performing detection processing using a low-resolution image, it is possible to detect the presence of the detection target object and its outline position at a high speed, and then use a higher-resolution image. By performing the detection process, the position of the detection target object can be specified with higher accuracy.

  Here, the detection principle that is the basis of the detection method in the detection apparatus 1 will be described with reference to FIGS. 6 and 7.

  FIG. 6 is a schematic diagram illustrating the principle of detection processing for detecting an object having a plurality of sizes using a pyramid image as described above. In FIG. 6, a plurality of size detection (multiple size detection) pyramid images PE1 are created by reducing the original image FE1, and search or collation (detection) is performed using the detection template TE1 for each created hierarchical image. FIG. 2 shows a detection method in which an object having a plurality of sizes existing in the original image FE1 can be detected using one detection template TE1.

  Here, it will be described in detail that an object having a plurality of sizes existing in the original image FE1 can be detected by one detection template TE1. In the detection process executed between the original image FE1 and the detection template TE1, an object having a size corresponding to the detection template TE1 existing in the original image FE1 is detected. On the other hand, in the detection process between the image FE2 obtained by reducing the original image FE1 and the detection template TE1, an object having a size corresponding to the detection template TE1 existing in the image FE2 is detected. An object having a size detected between the image FE2 and the detection template TE1 is an object larger than the size corresponding to the detection template TE1 when converted to the original image FE1. That is, in the detection process executed between the image FE2 obtained by reducing the original image FE1 and the detection template TE1, an object larger than the object having a size corresponding to the detection template TE1 is detected. Further, since the size of the object detected in this way is determined by the reduction ratio between the original image FE1 and the reduced image, an object having a plurality of sizes is changed to one detection template TE1 by changing the reduction ratio. It becomes possible to detect using.

  FIG. 7 is a schematic diagram showing the principle of detection processing for detecting an object of a predetermined size at high speed as described above. Specifically, as shown in FIG. 7, first, the detection information TE1 (specifically, the templates TE1, TE2, TE3, and the detection information TE1 are reduced by reducing the detection template TE1 and the detection target image KE at the reduction ratio β. ) And a high-speed detection pyramid image PS1 (specifically, images KE, KS2, and KS3). Then, an approximate position of the object is specified by executing an initial detection process using the generated low-resolution image KS3 and template TE3, and further, based on the information on the specified approximate position, a high-resolution image ( More accurate detection is performed using KS2, KE) and templates TE2, TE1 corresponding to the high resolution. As a result, an object present in the image KE can be detected at high speed.

  Then, by combining the above-described two detection processes, that is, for each hierarchical image in the pyramid image PE1 shown in FIG. 6, by using the high-speed detection method shown in FIG. 7, it exists in the original image FE1. It is possible to detect an object of a plurality of sizes at a high speed. FIG. 8 shows a detection process according to a comparative example using such a detection principle.

  On the other hand, the above-described detection process of FIG. 5 makes it possible to further increase the speed while using the detection principle as shown in FIG. Specifically, in the detection method executed by the detection apparatus 1 shown in FIG. 5 described above, a pyramid image MP1 is created by partially sharing the plurality of pyramid images PE1 and the high-speed pyramid image PS1 in FIG. By using this for the detection process, the pyramid image can be created more efficiently than the detection method shown in FIG. In other words, the detection device 1 can quickly detect a plurality of sizes of objects present in an image, and further, the detection processing time can be shortened by improving the efficiency of creating a pyramid image used for detection. Yes.

<Operation>
Below, each function of the detection apparatus 1 mentioned above is demonstrated in detail. Here, the case where the detection target object is a face will be described. However, the idea of the present invention can be applied if detection is performed assuming the size of the object on the target image. The object is not limited to the face.

  FIG. 9 is a flowchart showing the overall operation of the detection apparatus 1.

  First, after a detection start instruction (step S21) by an operation of the input unit 6 or the like performed by a user (user), detection is performed from a detection object template (hereinafter also referred to as “detection template”) which is reference information used for detection. The use information IF is set (step S22), and the hierarchical image group MP is created from the target image (step S23).

  Next, based on the information set in step S22, a search window area for performing raster scan on a predetermined image in the hierarchical image group MP is determined (step S24), and matching ( A collation process is performed (step S25). Further, it is determined whether or not all areas on the predetermined image have been searched (step S26). When it is determined that the search has been completed, the detection process is ended (step S27), and it is determined that the search has not been completed. Then, the process from step S24 to step S26 is repeatedly performed.

  Hereinafter, each process executed in steps S21 to S25 will be described in detail. FIG. 10 is a flowchart showing the information setting process. FIG. 11 is a flowchart showing pyramid image creation processing. FIG. 12 is a flowchart showing search window determination processing. FIG. 13 is a flowchart showing the matching process.

<Start face detection processing>
Before starting the face detection process, the user (user) designates the size of the face detected by the detection device 1. Specifically, the user (user) operates the input unit 6 and the like on the operation screen Q shown in FIG. 4 to specify the size of the face detected by the detection device 1. For example, when a numerical value “1” is input in the field R1 of the small window W1 in the operation screen Q and a numerical value “6” is input in the field R2, the face size corresponding to the input numerical value is displayed at the top of each field. Is done. In the column R1, the minimum size MN (minimum size) to be detected is specified, and in the column R2, the maximum size MX (maximum size) to be detected is specified. The user (user) compares the size of the face appearing in the image PI with the size of the face to be detected, and determines the size of the face to be detected. The determination is performed by selecting and pressing the execution button BT with the input unit 6 or the like in a state where numerical values are input in the respective columns. When the execution button BT is pressed in a state where the numerical value “1” and the numerical value “6” are input in the column R1 and the column R2, respectively, the size SZ1 corresponding to the numerical value “1” corresponds to the numerical value “6”. Faces of six types of sizes (SZ1 to SZ6) up to the size SZ6 to be performed are designated as detection targets (number of sizes M = 6), and detection processing is started, and information setting processing (step S22) ).

<Information setting process>
FIG. 14 is a diagram schematically illustrating the information setting process.

  In this embodiment, as in FIG. 7, when detecting a face of each size, comparison processing is performed using two-stage reduced images. Therefore, in addition to the detection template TP1, two templates TP2 and TP3 are created by sequentially reducing the template TP1 using the reduction ratio β.

  Specifically, as shown in FIG. 10, when the information setting process is started (step S31), the detection information IF is detected from the face detection template TP1 that is the reference information for detection held in the storage unit 3. (IF2) is set (step S32). Specifically, a reduced face template TP2 reduced by the reduction rate β = 1/2 is created from the face detection template TP1, and a re-reduction template TP3 is created from the reduced face template TP2 by the reduction rate β = 1/2. Is done.

  Here, the reduction ratio β is not limited to the above-mentioned ½, and may be various values, but is a numerical value satisfying the reduction ratio β = 1 / K (K is an integer of 2 or more). Preferably there is. In the present application, the “reduction ratio” of an image is represented by a magnification per side of the image. Image reduction can be done with existing interpolation methods such as bi-linear interpolation, nearest neighbor method, or bi-cubic convolution depending on processing time limitations and required image quality. It can be done using the method.

  Furthermore, in this embodiment, instead of directly comparing the reduced image template TP3 and the reduced image, by comparing the feature amount extracted from the reduced image template TP3 with the feature amount extracted from the reduced image, The case where the comparison process using a template is performed is illustrated. Therefore, in step S32, a feature amount CH in which the ratio of the skin color area to the entire area is quantified is extracted from the re-reduction template TP3. However, the present invention is not limited to this, and comparison processing using a template may be performed by pattern matching processing that directly compares the re-reduced template TP3 and the reduced image in units of pixels. The amount CH extraction process is not necessary.

  Each template and feature amount CH created or extracted in this way is read into the RAM 3b as detection information IF (IF2) and used in matching processing described later.

  Note that the number of template reductions S to be executed on the basis of the face detection template TP1 is not limited to the number of times as described above (that is, twice), and is set to an appropriate number depending on the relationship with the detection processing time, detection accuracy, and the like. It is possible to change. The feature amount CH is not limited to being extracted from the re-reduction template TP3, and the feature amount may be extracted from the reduction template TP2 or the face detection template TP1 and used for matching processing. Furthermore, here, as the extracted feature quantity CH, a value obtained by quantifying the ratio of the skin color area to the entire area of the re-reduction template TP3 as described above is illustrated, but the present invention is not limited to this, and the template is not limited thereto. It may be one obtained by extracting other features of the. Further, information obtained by parameterizing a template using principal component analysis (PCA) or the like may be used as a feature amount.

  When the detection information IF (IF2) is set, the information setting process ends (step S33), and the process proceeds to a pyramid image creation process (step S23).

<Pyramid image creation processing>
FIG. 15 is a diagram illustrating reference image creation in the pyramid image creation processing. FIG. 16 is a diagram illustrating reduced image creation in the pyramid image creation processing. FIG. 17 is a completed view of the pyramid image. FIG. 18 is a diagram illustrating image types in a pyramid image. FIG. 19 is a diagram illustrating a reduction ratio between images in a pyramid image.

  As shown in FIG. 11, when the pyramid image creation process is started (step S41), the number L of layers to be created (hereinafter also referred to as “total number of layers”) is determined (step S42). Specifically, it is determined by the total number of layers L = M + S × N derived from the relationship among the number M of face sizes to be detected, the number of template reductions S, and a reference image number N described later. In the present embodiment, a case will be described where the number M of face sizes to be detected is M = 6, the number of template reductions S = 2, and the number of reference images N = M−2 = 4. Therefore, in the present embodiment, the total number of hierarchies created is L = 14 hierarchies.

  Next, a first layer image F1 used to detect the minimum face size designated at the start of the detection process (step S21) is created (step S43). Specifically, it is created from the target image F0 (original image) using a predetermined magnification γ (see FIG. 15). The predetermined magnification γ is determined by a ratio between the face size LA1 of the face detection template TP1 and the minimum face size MN (minimum size) designated by the user (user). For example, when the designated size MN is twice the size LA1 of the template TP1, the magnification γ = 1/2, and the designated size MN and the size LA1 of the template TP1 are equal. In this case, the magnification γ = 1, that is, equal magnification.

As described above, when the first layer image F1 is created in step S43, (N-1), that is, three middle layer reference images G2 are created in the middle layer reference image creation step (step S44). (See FIG. 15). Specifically, first, an image F2 of the second hierarchy reduced from the image F1 of the first hierarchy by the reduction ratio α is created. Further, an image F3 in the third hierarchy is created from the image F2 in the second hierarchy with a reduction ratio α, and thereafter images up to the Nth hierarchy are sequentially created with the reduction ratio α. That is, when the reference image number N = 4, the reference image is created using the reduction ratio α up to the fourth-layer image F4. The first layer image F1 created in step S43 and the intermediate layer reference image G2 (images F2 to F4) created in step S44 constitute a reference image G3 (see FIG. 18). These reference images G3 are images that serve as a reference in creating a reduced image, as will be described below. The images F1 to F4, which are constituent elements of the reference image G3, have different magnifications relative to the target image F0. Specifically, the images F1 to F4 are γ times, γ × α times, γ × α ² times, and γ × α ³ times the target image F0, respectively.

When the intermediate layer reference image creation step (step S44) ends, the process proceeds to a reduced image creation step (step S45). In step S45, first, an image F5 corresponding to the (N + 1) th layer, that is, the fifth layer is created based on the image F1 of the first layer. Specifically, the fifth layer image F5 is created by reducing the first layer image F1 by the reduction ratio β (see FIG. 16). Here, the relationship β = α ^N is established among the reference image number N, the reduction rate α, and the reduction rate β. That is, in the case of the present embodiment (reference image number N = 4, reduction rate β = 1/2), the reduction rate α is automatically determined by the above relationship, and the reduction rate α≈1 / 1.99. In this embodiment, the reference image number N = 4. However, the reference image number N is an arbitrary number as long as it is an integer equal to or larger than 2 and less than the number M of face sizes to be detected.

  When the creation of the reduced image in step S45 is completed, the total number of hierarchies L determined in step S42 is compared with the number of hierarchies Fk of images already created in step S45 (step S46). If it is determined by this comparison that layer images corresponding to the total number of layers L have been created, the pyramid image creation process is terminated (step S47), and if the number of created layers Fk is less than the total number of layers L, the number of created layers 1 is added to Fk (step S48). For example, when Fk = 5, the process proceeds to step S48 and Fk = 6.

  Thereafter, the reduced image processing (step S45) is executed again for the new upper hierarchy 6. The image F6 of the sixth hierarchy is created by using the image F2 of the second hierarchy as a reference, and the image F2 of the second hierarchy is reduced by the reduction ratio β (see FIG. 16). Thereafter, similarly, the image F3 in the third hierarchy is used as a reference for the creation of the image F7 in the seventh hierarchy, and the image F4 in the fourth hierarchy is used as the reference for the creation of the image F8 in the eighth hierarchy. . Further, the images after the ninth layer are similarly created based on the images four layers below the creation layer, and the steps S45 to S48 are repeatedly executed until the number of layers Fk = 14 (see FIG. 17).

  As shown in FIG. 17, when the image for the total number of hierarchies L, that is, the image F14 of the 14th hierarchy is created by the above-described repetition process, the pyramid image creation process ends (step S47) and the search window is determined. The process proceeds to processing (step S24).

  When the above-described pyramid image creation processing (step S23) is summarized with reference to FIG. 18, a first-layer image F1 corresponding to the reference image G1 is created in step S43, and based on the reference image G1 (image F1) in step S44. Thus, the intermediate layer reference image G2 is created with the reduction ratio α. Next, in the iterative processing from step S45 to step S48, a reduced image GM1 (images F5 to F8) is created from the reference image G3 (images F1 to F4) with a reduction ratio β, and is reproduced based on the created reduced image GM1. The reduced image GM2 (images F9 to F12) is created with the reduction ratio β, and the re-reduced image GM3 (images F13 and F14) is created based on the re-reduced image GM2.

  All images created by this iterative process are created using a reduction ratio β = 1 / K (K is an integer equal to or greater than 2), so a pyramid image is created rather than using a numerical value such as the reduction ratio α. Can be speeded up.

  Note that, in the pyramid image MP, the degree of reduction increases as the hierarchy becomes higher (in other words, the magnification with respect to the original image F0 decreases), so the resolution becomes lower as the hierarchy becomes higher. In the present embodiment, since the total number of layers to be created is L = 14, the images corresponding to the re-reduced image GM3 include the ninth-layer image F9 and the tenth-layer image F10 belonging to the re-reduced image GM2. The images are based on the 13th layer image F13 and the 14th layer image F14 created based on the above. Assuming that the total number of layers L = 13 (M = 5, S = 2, N = 4), an image corresponding to the re-reduced image GM3 is an image F9 of the ninth layer belonging to the re-reduced image GM2. Only the image 13 of the thirteenth layer created based on this is obtained.

  In FIG. 18, the multiple-size detection pyramid image H1 used for detecting an object having a plurality of sizes in the target image is configured to include six images F1 to F6 having different magnifications with respect to the target image. Yes. Further, the high-speed detection pyramid image H2 used for detecting each of a plurality of objects in the target image at high speed is composed of ten images F5 to F14. Specifically, the high-speed detection pyramid image H2 includes six images F5 to F10 corresponding to a state in which the six images F1 to F6 in the multiple-size detection pyramid image H1 are respectively reduced at a predetermined magnification β. . The high-speed detection pyramid image H2 includes six images F9 to F14 corresponding to a state in which the six images F5 to F10 are further reduced by a predetermined magnification β.

  Here, the multiple-size detection pyramid image H1 and the high-speed detection pyramid image H2 are created by the above-described method or the like, so that the magnifications of the pyramid images H1 and H2 with respect to the target image are mutually equal. The partial images (specifically, images F5 and F6) are created in a common state so as to be the same. Thereby, the time required for creating the pyramid image can be shortened. This will be described later.

Note that in the pyramid image creation process (step S23) when the reference image number N = 4 and the reduction rate β = 1/2, the reduction rate α is strictly reduced from the relationship of the reduction rate β = α ^N. In this embodiment, the rational number α = 1 / 1.99 is used as the value satisfying the relationship of the reduction ratio β = α ^N. Yes. For this reason, α ^N does not completely match the reduction ratio β. However, this discrepancy is within an allowable error range and does not cause inconvenience in the pyramid image creation processing. This is because, as shown in FIG. 19, even when the uniform reduction ratio α = 1 / 1.99 is used when creating the intermediate layer reference image G2, the fifth layer image F5 is originally in the first layer. This is because the image is created from the image F1 with the reduction ratio β, and similarly, all the images in the sixth and subsequent layers are also created using the reduction ratio β, so that the pyramid image creation is not affected.

<Search window decision processing>
Next, before the search window determination process (step S24) is described, the search method in the detection apparatus 1 will be described in detail with reference to FIGS.

  FIG. 20 is a diagram illustrating a correspondence relationship between the detection information IF (IF2) and the pyramid image MP (MP2) when detecting the face having the minimum size. FIG. 21 is a diagram illustrating a target hierarchy in a pyramid image corresponding to each size to be detected. FIG. 22 is a diagram illustrating a hierarchy used by each attention hierarchy for detection in the pyramid image.

  As shown in FIG. 20, in the case of detecting the face (size SZ1) having the minimum size designated by the user (user), the first layer image F1, the face detection template TP1, and the fifth The image F5 and the reduced face template TP2 in the hierarchy, and the feature amount CH extracted from the image F9 in the ninth hierarchy and the re-reduced template TP3 are respectively detected by comparison.

  Here, a hierarchy serving as a reference in comparison detection is referred to as an attention hierarchy. That is, the target hierarchy when detecting the size SZ1 is the first hierarchy. For example, as shown in FIG. 21, when a face of size SZ2 is detected, the attention layer is the second layer. In this case, the layer to be compared with the reduced face template TP2 is the sixth layer, and the layer to be compared with the feature amount CH is the tenth layer. Further, in the case of detecting faces of various sizes from size SZ3 to size SZ6, the third to sixth layers are the attention layers, respectively. The correspondence relationship described above is shown in detail in FIG. For example, when a face of size SZ5 is detected, the target layer (corresponding to the part marked with ◎ in FIG. 22) is the fifth layer and the layer to be compared with the reduced face template TP2 (part marked with ◯) Corresponds to the ninth layer, and the layer to be compared with the feature amount CH (corresponding to the part marked with Δ) is the thirteenth layer.

  As shown in FIG. 22, when the target layer is the first layer image F1 to the sixth layer image F6, the number of layers of the pyramid image required for the detection process is from the first layer image F1 to the first layer. There are 14 layers up to the image F14 of 14 layers. This coincides with the total number of pyramid images created in step S42.

  Here, as in the present embodiment, a detection method is employed in which a multi-size detection pyramid image PE1 and a high-speed detection pyramid image PS1 as shown in the comparative example of FIG. Assume that six face sizes are detected at high speed. In this case, images of six layers corresponding to the target layer are required for the pyramid images for size detection, and further, two layers for each layer of the target layer are included in the pyramid image for high speed detection, that is, 6 × 2 = 12 layers. Need an image. That is, the total total number Ln of pyramid images required for detection is an image for 12 + 6 = 18 layers. Further, in the detection method shown in the comparative example of FIG. 8, the total number Ln of pyramid images required for detection is Ln = M + M × S from the relationship between the number M of face sizes to be detected and the number of template reductions S. Desired.

  As described above, the reason why the number L of pyramid images created by the detection apparatus 1 is smaller than that of the detection method shown in FIG. 8 is that the detection apparatus 1 partially uses a plurality of size detection pyramid images and a high-speed detection pyramid image. This is because the efficiency of pyramid image creation is realized by creating a common pyramid image MP (MP2). More specifically, as shown in the fifth and sixth hierarchies C1 or the ninth and tenth hierarchies C2 in FIG. 22, the detection apparatus 1 uses the image once used in the detection process as another image. It is possible to reduce the total number of images to be created by creating pyramid images so that they can be reused even when detecting a size face. Specifically, in this case, (reference image number N) <(number M of face sizes to be detected) holds, so (number of created images L in the present embodiment) <(in the comparative example of FIG. 8). The number of created images Ln).

  The fifth-layer image F5 and the sixth-layer image F6 are images belonging to both pyramid images (image groups) of the multiple-size detection pyramid image and the high-speed detection pyramid image, and the magnifications with respect to the target image F0 are respectively set. Are identical to each other. Therefore, the pyramid image MP (MP2) is a pyramid image in which a plurality of size detection pyramid images and a high-speed detection pyramid image are partially shared.

  As described above, the detection device 1 can detect a plurality of objects having a plurality of sizes in an image at a high speed, and further reduce the number of creation by using the pyramid image used for detection, thereby creating a pyramid image. By shortening the time required, the detection processing is further speeded up.

  Further, the detection apparatus 1 creates the fifth layer image F5 corresponding to the target layer by reducing the image F1 from the first layer with the reduction rate β, and similarly, the sixth layer image F6 is also the second layer image F6. It is created by reducing the image F2 in the hierarchy with the reduction ratio β. In this way, by using the reduction ratio β = 1 / K (K is an integer of 2 or more), the interpolation is more complicated than when using the reduction ratio α = 1 / 1.19 or the like. Since it is not necessary to perform processing, it is possible to shorten the processing time for creating a reduced image.

  On the other hand, the detection method shown in the comparative example of FIG. 8 is not a creation method that effectively uses the reduction rate β when creating an image for the target layer.

  In other words, the detection apparatus 1 can shorten the processing time required to create the fifth-layer image F5 and the sixth-layer image F6, and realizes a high-speed detection process.

  Next, the order of comparison detection will be described. The comparison detection is first performed between the most reduced images, that is, between the image F9 in the ninth layer and the re-reduction template TP3 (CP1). More specifically, the feature amount (skin color frequency) extracted from the image F9 is compared with the feature amount CH (skin color frequency) extracted from the re-reduction template TP3. This first comparison detection CP1 is performed in the entire region of the image F9. When a face having a size to be detected is detected by the first comparison detection CP1, the image F5 in the fifth hierarchy is detected based on the detected position. And the reduced face template TP2 are compared and detected (CP2). Further, if further detection is performed in the second comparison detection CP2, comparison detection is performed between the image F1 on the first layer and the face detection template TP1 based on the detected position (CP3). When the detection is performed also in the third comparison detection CP3, the face of the detection target size is output as being present in the target image.

  As described above, the processing time required for detection can be shortened by first performing comparative detection on the entire reduced image and specifying the approximate position of the detection target object.

  Next, the search window determination process will be described with reference to FIGS. 20, 23, and 24.

  FIG. 23 is a diagram illustrating a detection processing procedure when the target layer is the first layer. FIG. 24 is a diagram illustrating search window shift processing. In FIG. 24, the upper left corner of the image F9 is the reference position P (1, 1), and the lower right corner is P (Xm, Ym). Further, the position of each search window is represented by the coordinates of the upper left corner of each search window.

  As shown in FIG. 12, when the search window determination process is started (step S51), a search window for executing a matching process described later is determined (step S52).

  FIG. 23 illustrates the detection method when the target layer is the first layer, as in FIG. 20. As described above, the process CP1 that is executed first is between the most reduced images, that is, the first layer. This is performed between the image F9 of the nine layers and the feature amount CH extracted from the re-reduction template TP3. More specifically, the search window IW having an area having the same size as the re-reduction template TP3 is set at the upper left corner of the image F9. When the setting of the search window IW is completed (step S53), the process proceeds to a matching process (step S25) described later. In the matching process (step S25), the matching process using the feature amount CH is performed in the set search window IW. After the matching process is completed, the search window determination process is started again (step S51), and a location shifted by one pixel in the x-axis direction from the previous matching process position is set as a new search window IW (step S52). The processes from step S51 to step S53 are executed until the search window IW shifts and covers all areas of the image F9 in the ninth hierarchy.

  The shift process of the search window IW will be described in more detail with reference to FIG. The search window IW is shifted pixel by pixel in the x-axis direction from the initial position P (1, 1) at the upper left corner in the image F9 in the ninth layer to the upper right corner position P (Xm, 1) (see FIG. 24A). ). When the search window IW reaches the upper right corner position P (Xm, 1), the search window IW is set at a position P (1, 2) shifted by one pixel in the y-axis direction from the initial position P (1, 1). Thereafter, similarly, the same scanning is performed from the position P (1, 2) in the x direction, and the search window IW finally moves to the position P (Xm, Ym). That is, the scanning process of the search window IW is performed with the x direction as the main scanning direction and the y direction as the sub scanning direction. Note that the shift width is not limited to one pixel as described above, and may be changed to an appropriate value depending on the relationship with the detection processing time, the detection accuracy, and the like. If the shift width is increased, the detection accuracy becomes coarse, but the detection processing time is shortened.

<Matching process>
Hereinafter, the matching process (step S25) will be described with reference to FIGS.

  FIG. 25 is a diagram illustrating one-dimensionally the shift state in the low-resolution image and the shift state in the high-resolution image.

  FIG. 26 is a diagram two-dimensionally showing the shift state in the high resolution image.

  As shown in FIG. 13, when the matching process is started (step S61), the skin color area determination process is executed in the search window IW (step S62). The skin color area determination is a process corresponding to the first comparison detection CP1 (see FIG. 23) described above, and the ratio of the skin color pixels to the entire area extracted from the reduced template TP3, that is, the area occupied by the feature amount CH and the search window IW The ratio of the skin color pixels inside is compared. For example, the three primary color information of red (R), green (G), and blue (B) included in the search window IW is converted to a color system of hue (H), color (S), and brightness (V), and color space conversion is performed. Judgment is made based on whether or not the value of the hue (H) is within a range indicating a preset skin color, the number of skin color pixels in the search window IW region is calculated, and the number of skin color pixels is determined from the number of skin color pixels of the feature amount CH If the threshold is equal to or greater than the threshold value, it may be determined that the skin color region has been detected.

  If the skin color area is not detected in the skin color area determination process (step S62) (step S63), the matching process is terminated (step S64), and the process proceeds to the process of determining the next search window IW through step S26 (FIG. 9). When the skin color area is detected (step S63), the comparison detection CP2 is performed between the above-described fifth layer image F5 and the reduced face template TP2 based on the detected position (step S65).

  Furthermore, if it is not detected in step S65 that there is a face of the size to be detected (step S66), the matching process is terminated (step S67), and the process proceeds to step S26 to determine the next search window IW. Transition (see FIG. 9). When it is detected that a face of the size to be detected exists (step S66), based on the detected position, comparison detection CP3 by template matching is performed between the image F1 on the first layer and the face detection template TP1. Performed (step S68).

  As described above, in the first comparison detection CP1, by performing skin color region determination using the feature amount CH, template matching of the search window IW having no skin color region can be omitted, and detection processing is performed at high speed. be able to. That is, it is possible to narrow down an area where a face is likely to exist from the entire area by a comparison process using a low resolution image. In other words, it is possible to exclude in advance an area that is determined to have no face in the comparison process using the skin color area from the comparison process using the high-resolution image.

  When step S68 ends, the matching process ends (step S69), and the process proceeds to step S26. When a face having a size to be detected is detected in step S68, detection information (for example, detection position or size) is stored in the RAM 3b as output information, and is output after completion of all the processing (step S27). The

  Hereinafter, the matching process will be described in detail with specific examples.

  For example, if a skin color region is detected at the position P (v, w) in the image F9 of the ninth hierarchy in FIG. 23 in step S62, the position P (v, w) in the image F5 of the fifth hierarchy is used as the basis. Thus, comparison detection CP2 is performed by template matching with the reduced face template TP2. Specifically, as shown in FIG. 25, in comparison detection CP1, the search window IW is shifted to position Pv-2, position Pv-1, and further to position Pv, and in skin color region determination at position Pv (step S62). When the skin color area is detected, the process proceeds to comparison detection CP2 (step S65) between the image F5 of the fifth layer and the reduced face template TP2, and template matching is performed at two points of the position Pv and the position Pv + 0.5. Executed.

  Here, since the image F9 used in the comparison detection CP1 is an image obtained by halving the image F5 used in the comparison detection CP2, the shift width of one pixel of the search window IW in the image F9 is 0 in the image F5. This corresponds to 5 pixels.

  Therefore, if the shift width of the template matching performed in the comparison detection CP2 is 1 pixel, the position where the template matching is performed in the comparison detection CP2 is the above-described two points shown in FIG. In the two-dimensional executed, there are four points P (v, w), P (v + 0.5, w), P (v, w + 0.5) and P (v + 0.5, w + 0.5) in the image F9. The corresponding position. FIG. 26 shows the state of shifting at the four points described above, and corresponds to a diagram in which the region PR2 in FIG. 23 is enlarged and displayed. In the image F5, template matching is performed at a position PA corresponding to P (v, w) in the image F9, and then a position PB (P (v + 0.P in the image F9) shifted by one pixel from the position PA in the x-axis direction. 5), template matching is performed (see FIG. 26A). Further, template matching is performed at a position PC (corresponding to P (v, w + 0.5) in the image F9) shifted by one pixel from the position PA in the y-axis direction. Template matching is performed at the position PD shifted by the minute (corresponding to P (v + 0.5, w + 0.5) in the image F9) (see FIG. 26B).

  In template matching, a correlation value between images to be compared (for example, a comparison region between the reduced face template TP2 and the fifth layer image F5) is calculated, and the calculated correlation value is compared with a preset threshold value. Thus, it is determined whether or not a face having a size to be detected exists. For example, when the detection template has 100 × 100 pixels, the absolute difference value of the pixel signal (here, color difference information) is obtained between 10,000 corresponding two pixels, and 10,000 pixels obtained thereby are obtained. The average value of the absolute differences can be used as the correlation value. Such a correlation value is a small value when the corresponding pixels include the same part of the subject, and a large value when the corresponding pixels are different parts.

  In addition, although the shift width of the template matching performed in the image F5 is 1 pixel, it is not limited to this. For example, the shift width sb executed in the image BB is within a range in which the predetermined shift width sa adopted in the predetermined reduced image AA is not larger than the shift width sa2 when converted to the image BB that is the original of the reduced image AA. Is used, the detection performed on the image BB can be performed with the same or high accuracy as the detection performed on the reduced image AA. That is, if the shift width of the search window IW in the image F9 is 1 pixel, the shift width of 2 pixels or less may be set in the image F5.

  As described above, the skin color area is detected in step S62, and the matching process executed in step S65 based on the detection position has been described in detail. However, there may be a case where it is detected in step S65 that a face having a size to be detected exists. Similar processing is performed. That is, in step S68, comparison detection CP3 by template matching between the image F1 of the first hierarchy and the face detection template TP1 is executed at the four points obtained from the position detected in step S65. For example, if it is detected at the position PW2 (PC) in the fifth-layer image F5 in step S65, the region in FIG. 23 obtained based on the position PW1 in the first-layer image F1 corresponding to the position PW2. Comparison detection CP3 by template matching is performed at four points indicated by PR1.

  As described above, in an operation for detecting an object having a size to be detected in an image of a certain target hierarchy (hereinafter, also simply referred to as a “target image”) (that is, an object detection operation of a certain size), the target image and the target The target object is obtained by performing comparison processing using a template for images determined in order from low-resolution images among images obtained by reducing images at a predetermined magnification β (images in the high-speed detection pyramid image H2). Is detected. For example, when the face detection process of the minimum size SZ1 is performed, a comparison process (comparison detection CP1) using the low resolution image F9 and the template TP3 is first performed, and then the medium resolution image F5 and the template are performed. Comparison processing (comparison detection CP2) using TP2 is performed, and finally comparison processing (comparison detection CP3) using the high-resolution image F1 and template TP1 is performed. As a result, whether or not a face having the minimum size SZ1 is present in the target image, and if present, its presence position is detected. When face detection processing with a larger size SZ2 is performed, first, comparison processing (comparison detection CP1) using the low-resolution image F10 and the template TP3 is performed, and then the medium-resolution image F6 and Comparison processing using the template TP2 (comparison detection CP2) is performed, and finally comparison processing using the high-resolution image F2 and the template TP1 (comparison detection CP3) is performed. As a result, the position of the face having the size SZ2 is detected. The same applies to the other sizes SZ3 to SZ6.

  In this manner, an object having a size to be detected in the target image is detected by the comparison process using the target image selected from the pyramid image H1 for multiple size detection (see FIG. 18) and the template. Operation is performed.

  Then, by repeating such an operation for a plurality of images of interest, a plurality (6 in the above example) of objects SZ1 to SZ6 are detected.

<Others>
Although the embodiments of the present invention have been described above, the present invention is not limited to the contents described above.

  For example, in the above-described embodiment, skin color region determination is performed in the comparison detection CP1, and template matching is performed in the comparison detection CP2 and CP3, and the detection process is performed. However, the present invention is not limited to this. Specifically, template matching may be executed in all of the comparison detections CP1 to CP3, and a feature amount such as skin color region determination is used prior to template matching executed in all of the comparison detections CP1 to CP3. A determination process may be executed. As described above, by executing the simple determination process prior to the template matching, the matching process for the search window having no skin color region can be omitted, and the detection process can be executed at high speed.

In the above embodiment, the reduction rate α is a value that satisfies the relationship of reduction rate β = α ^N , but is not limited to this. Specifically, the reduction rate α is a value used to create the reference image GM3. If the reduction rate β is a constant value, it does not affect the pyramid image creation. Therefore, an arbitrary value that does not satisfy the relationship of the reduction rate β = α ^N can be set as the reduction rate α.

  Further, in the above-described embodiment, the face size specified by the user (user) on the operation screen Q shown in FIG. 4 is a mode in which the user selects a face size set in advance. However, it is not limited to this. Specifically, since the reduction ratio α can be set to an arbitrary value as described above, the predetermined magnification γ used when creating the first layer image F1 from the target image F0 (original image) is also arbitrary. If the user can set the face size itself, the user (user) can freely set the face size. That is, in creating the reference image GM3, a predetermined magnification γ and a reduction rate α are determined so that an image of a target layer for detecting the size of the face designated by the user (user) can be created. You may make it an aspect.

It is a figure which shows the outline | summary of the detection apparatus which concerns on embodiment of this invention. It is a block diagram which shows the various functions with which a detection apparatus is provided. It is a figure which shows the image which has several persons as a to-be-photographed object. It is a figure which shows the operation screen at the time of program execution. It is the figure which showed typically an example of the detection process performed in a detection apparatus. It is a schematic diagram of the detection process which detects the object of the some magnitude | size which exists in an original image using a pyramid image. It is a schematic diagram of the detection process which detects the predetermined object which exists in an image at high speed. It is a schematic diagram of the detection process which concerns on a comparative example. It is a flowchart which shows the whole operation | movement of a detection apparatus. It is a flowchart which shows an information setting process. It is a flowchart which shows a pyramid image creation process. It is a flowchart which shows a search window determination process. It is a flowchart which shows a matching process. It is the figure which showed the information setting process typically. It is a figure which shows the reference | standard image creation in a pyramid image creation process. It is a figure which shows the reduction image creation in a pyramid image creation process. It is a completion drawing of a pyramid image. It is a figure which shows the image classification in a pyramid image. It is a figure which shows the reduction magnification between each image in a pyramid image. It is a figure which shows the correspondence of the information for a detection in the case of detecting the face of the minimum magnitude | size, and a pyramid image. It is a figure which shows the attention hierarchy in the pyramid image corresponding to each magnitude | size to detect. It is the figure which showed the hierarchy which each attention hierarchy uses for a detection in a pyramid image. It is a figure which shows the detection process procedure at the time of making an attention hierarchy into the 1st hierarchy. It is a figure which shows the shift process of a search window. It is a figure which shows the mode of the shift in a low resolution image, and the mode of the shift in a high resolution image in one dimension. It is a figure which shows the mode of the shift in a high resolution image in two dimensions.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 Detection apparatus 3 Memory | storage part 3a Hard disk drive (HDD)
3b RAM
DESCRIPTION OF SYMBOLS 5 Display part 6 Input part 6a Keyboard 6b Mouse 11 Information setting part 12 Pyramid image creation part 13 Search window determination part 14 Matching process part

Claims (18)

A detection device for detecting an object present in a target image,
A first image group used to detect objects of a plurality of sizes in the target image, the first image group having a plurality of images with different magnifications relative to the target image; A plurality of images corresponding to a state in which each of the plurality of images in the first image group is reduced at a predetermined magnification. A hierarchical image group creating means for creating a hierarchical image group including a second image group having a plurality of images;
Object detection means for detecting the object using the hierarchical image group;
With
In the first image group and the second image group, a part of the images is made common so that the magnifications of the part of the images in the image group with respect to the target image are the same. A detection device produced by the method described above. The detection device according to claim 1,
The hierarchical image group creating means creates a plurality of reference images having different magnifications with respect to the target image and creates a plurality of reduced images obtained by further reducing each of the plurality of reference images at a predetermined magnification. A hierarchical image group including the reference image and the plurality of reduced images,
A part of the plurality of reduced images is created in common as an image belonging to both the first image group and the second image group. The detection device according to claim 2,
The hierarchical image group creating means includes
Determining means for determining the first reduction rate α and the second reduction rate β so as to satisfy a relationship of β = α ^N (N is an integer of 2 or more);
A first reference image is created based on the target image, and the first reference image is sequentially reduced at an inter-image reduction ratio of the first reduction ratio α, so that (N−1) intermediate layer references Means for creating the plurality of reference images including the first reference image and the (N-1) intermediate layer reference images by creating an image;
Means for creating a plurality of reduced images by further reducing each of the plurality of reference images at the second reduction rate β;
A detection apparatus comprising: The detection device according to claim 3,
The hierarchical image group creating means includes
Means for creating at least one re-reduced image by further reducing one image or each of the plurality of reduced images at the second reduction rate β;
The detection apparatus further comprising: In the detection device according to claim 3 or 4,
The second reduction rate β is 1 / K (K is an integer equal to or greater than 2). The detection device according to any one of claims 1 to 5,
The object detection means includes
The object of the plurality of sizes is obtained by repeating an operation of detecting an object of a size to be detected in the target image by a comparison process using the target image selected from the first image group and the template. Detect
In the operation of detecting an object having a size corresponding to the target image, the target image and an image obtained by reducing the target image at the predetermined magnification and having a low resolution out of the second group of images. A detection apparatus characterized by efficiently detecting an object having a size to be detected in the target image by detecting an object in order from an image. A detection method for detecting an object present in a target image,
A first image group used to detect objects of a plurality of sizes in the target image, the first image group having a plurality of images with different magnifications relative to the target image; A plurality of images corresponding to a state in which each of the plurality of images in the first image group is reduced at a predetermined magnification. A hierarchical image group creating step of creating a hierarchical image group including a second image group having the images of:
An object detection step of detecting the object using the hierarchical image group;
Including
In the first image group and the second image group, a part of the images is made common so that the magnification of the part of each image group with respect to the target image is the same. The detection method characterized by being created by. The detection method according to claim 7,
The hierarchical image group creating step creates a plurality of reference images having different magnifications with respect to the target image and creating a plurality of reduced images obtained by further reducing each of the plurality of reference images at a predetermined magnification. A hierarchical image group including the reference image and the plurality of reduced images,
A part of the plurality of reduced images is created by being shared as an image belonging to both the first image group and the second image group. The detection method according to claim 8, wherein
The hierarchical image group creation step includes:
A determination step of determining the first reduction rate α and the second reduction rate β so as to satisfy a relationship of β = α ^N (N is an integer of 2 or more);
A first reference image is created based on the target image, and the first reference image is sequentially reduced at an inter-image reduction rate of the first reduction rate α, thereby (N−1) intermediate layer references. Creating the plurality of reference images including the first reference image and the (N-1) intermediate layer reference images by creating an image;
Creating a plurality of reduced images by further reducing each of the plurality of reference images at the second reduction rate β;
A detection method comprising: The detection method according to claim 9, wherein
The hierarchical image group creation step includes:
Creating at least one re-reduced image by further reducing one image or each of the plurality of reduced images at the second reduction rate β;
A detection method characterized by further comprising: The detection method according to claim 9 or 10,
The second reduction ratio β is 1 / K (K is an integer of 2 or more). The detection method according to any one of claims 7 to 11,
The object detection step includes
The object of the plurality of sizes is obtained by repeating an operation of detecting an object of a size to be detected in the target image by a comparison process using the target image selected from the first image group and the template. Detect
In the operation of detecting an object having a size corresponding to the target image, the target image and an image obtained by reducing the target image at the predetermined magnification and having a low resolution out of the second group of images. A detection method characterized by efficiently detecting an object having a size to be detected in the target image by detecting an object in order from an image. On the computer,
A first image group used for detecting an object of a plurality of sizes in the target image, the first image group having a plurality of images with different magnifications relative to the target image; and A plurality of second image groups used to detect each of the plurality of objects at a high speed, each of which corresponds to a state in which the plurality of images in the first image group are reduced at a predetermined magnification. A hierarchical image group creating step of creating a hierarchical image group including a second image group having an image;
An object detection step of detecting the object using the hierarchical image group;
Is a program for detecting an object present in the target image by executing
In the first image group and the second image group, a part of the images is made common so that the magnifications of the part of the images in the image group with respect to the target image are the same. A program characterized by being created in. The program according to claim 13, wherein
The hierarchical image group creating step creates a plurality of reference images having different magnifications with respect to the target image and creates a plurality of reduced images obtained by further reducing each of the plurality of reference images at a predetermined magnification. Creating the hierarchical image group including the reference image and the plurality of reduced images,
A part of the plurality of reduced images is created in common as an image belonging to both the first image group and the second image group. The program according to claim 14, wherein
The hierarchical image group creation step includes:
In the computer,
A determination step of determining the first reduction rate α and the second reduction rate β so as to satisfy a relationship of β = α ^N (N is an integer of 2 or more);
A first reference image is created based on the target image, and the first reference image is sequentially reduced at an inter-image reduction rate of the first reduction rate α, thereby (N−1) intermediate layer references. Creating the plurality of reference images including the first reference image and the (N-1) intermediate layer reference images by creating an image;
Creating a plurality of reduced images by further reducing each of the plurality of reference images at the second reduction rate β;
A program characterized by having executed. The program according to claim 15, wherein
The hierarchical image group creation step includes:
In the computer,
Creating at least one re-reduced image by further reducing one image or each of the plurality of reduced images at the second reduction rate β;
Is further executed. In the program according to claim 15 or 16,
The second reduction ratio β is 1 / K (K is an integer of 2 or more). The program according to any one of claims 13 to 17,
The object detection step includes
In the computer,
The object having the plurality of sizes is detected by repeating the operation of detecting the object having the size corresponding to the target image by the comparison process using the target image selected from the first image group and the template. And
In the operation of detecting an object having a size as a detection target in the target image, the target image and an image obtained by reducing the target image at the predetermined magnification, which is lower than the second image group. A program characterized in that an object having a size to be detected in the target image is efficiently detected by detecting objects in order from the resolution image.

Images