JP4795737B2 - Face detection method, apparatus, and program - Google Patents

Description

  The present invention relates to a face detection method and apparatus for detecting a face image from a target image, and a program therefor.

  Conventionally, the color distribution of a person's face area in a snapshot photographed by a digital camera is examined to correct the skin color, or a person in a digital image photographed by a digital video camera of a surveillance system is recognized. Has been done. In such a case, since it is necessary to detect a face region corresponding to a person's face in the digital image, various techniques for detecting a face in the digital image have been proposed so far. Among them, a method using a discriminator module (hereinafter simply referred to as a discriminator) generated by machine learning learning using sample images is known as a face detection method that is considered to be particularly excellent in detection accuracy and robustness. (For example, refer nonpatent literature 1, patent documents 1-3, etc.).

  This method includes a face sample image group composed of a plurality of different face sample images having substantially the same face orientation and vertical direction, and a non-face sample image group composed of a plurality of different non-face sample images that are known not to be faces. Is used to learn the characteristics of being a face, and to generate and prepare a discriminator that can determine whether an image is a face image in a predetermined orientation and vertical direction. A face on the detection target image is detected by sequentially cutting out partial images in a target image (hereinafter referred to as a detection target image) and determining whether or not the partial image is a face. It is a technique to do.

  In this method, since it is determined whether or not each of the sequentially extracted partial images is a face, the amount of processing becomes enormous if it is attempted to scrutinize from the beginning, and it takes time to detect the face. There is. Therefore, in order to increase the efficiency of the discrimination process in the method using this discriminator, first, a relatively rough face detection process (for example, thinning out the positions of partial images to be sequentially cut out) is performed. There is a method of extracting face candidates and then performing fine discrimination processing on the extracted image near the face candidates to determine whether or not the face is a true face.

  By the way, since the discriminator is generally learned using a sample image with relatively good image quality, it is basically made for an image with a good image quality. On the other hand, as the detection target image, images with various brightness and contrast of the shooting scene are assumed. Therefore, for example, when the detection target image is an image taken in a dark place, even if an attempt is made to search for a dark part of the eye and a bright part of the nose representing facial features in this image, the brightness of the image affects, Searching may be difficult.

  For this reason, normalization is performed so that the contrast of the image is set to a certain level as a pre-process for the image to be detected or discriminated so that the face can be detected even if the brightness and contrast of the detection target image are different. A technique for performing processing has been proposed. The following three methods are mainly proposed as a method for performing this normalization process.

The first method is a method of performing conversion according to a conversion curve (lookup table) that approximates the pixel value of the entire image of the detection target image to a value representing the logarithm of the luminance of the subject in the image. The second method is a method of converting the pixel values for each partial image cut out from the detection target image so that the degree of dispersion of the pixel values (luminance values) in the region is made constant. The third method scans the local area of a predetermined size within the area of the partial image cut out from the detection target image, and adjusts the degree of dispersion of the pixel values in the local area to a certain level. Is a method to convert.
"High-speed omnidirectional face detection", Shihong LAO et al., Image Recognition and Understanding Symposium (MIRU2004), July 2004, P.II-271-II-276 Japanese Patent Application No. 2003-316924 Japanese Patent Application No. 2003-316925 Japanese Patent Application No. 2003-316926

  However, the first method is characterized in that the processing result is short, but the processing result is easily affected by the oblique light and background in the detection target image. The third method has the characteristic that the processing result is oblique light in the detection target image. However, the second method has a feature that the processing time is long, and the second method has a large influence of the processing time, the oblique light in the detection target image on the processing result and the background. Both are characterized by being both average.

  In other words, both methods are pros and cons, and no method has been proposed that takes advantage of the benefits of each of these normalization processes. This is in order to achieve highly accurate and efficient face detection. It has become one of the obstacles.

  In view of the circumstances described above, the present invention aims to provide a face detection method and apparatus capable of highly accurate and efficient face detection, and a program therefor, by taking advantage of these normalization processes. To do.

  The face detection method of the present invention converts an input detection target image, which is a target for detecting a face, according to a conversion curve that approximates the pixel value of the entire image to a value representing the logarithm of the luminance of the subject in the image. An overall normalization step for performing an overall normalization process; and a feature amount relating to a distribution of pixel values of the image based on the detection target image subjected to the overall normalization process; and the detection using the feature amount A face image candidate detecting step for detecting a face image candidate in the target image, and a pixel value on the image for the partial image corresponding to the face image candidate in the detection target image subjected to the overall normalization process; A local normalization step for performing a local normalization process for converting the degree of dispersion of the pixel value for each local area of a predetermined size in the area so as to approach a predetermined level, and the part subjected to the local normalization process And determining a feature amount related to the distribution of pixel values of the image based on the image, and determining whether the face image candidate corresponding to the partial image is a face image using the feature amount. It is characterized by this.

  The face detection device of the present invention converts an input detection target image, which is a target for detecting a face, according to a conversion curve that approximates the pixel value of the entire image to a value representing the logarithm of the luminance of the subject in the image. An overall normalization unit for performing an overall normalization process; and a feature amount related to a distribution of pixel values of the image based on the detection target image subjected to the overall normalization process; and the detection using the feature amount Face image candidate detecting means for detecting a face image candidate in the target image, and a pixel value on the image for the partial image corresponding to the face image candidate in the detection target image subjected to the overall normalization process A local normalization unit for performing local normalization processing for converting the degree of dispersion of pixel values for each local region of a predetermined size in the region so as to approach a predetermined level; and a partial image subjected to the local normalization processing. Z And determining means for calculating a feature amount related to the distribution of pixel values of the image and determining whether the face image candidate corresponding to the partial image is a face image using the feature amount. It is a feature.

  The program of the present invention converts a computer according to a conversion curve that approximates the pixel value of the entire image to a value representing the logarithm of the luminance of the subject in the image for an input detection target image that is a target for detecting a face. An overall normalization means for performing the overall normalization process, and calculating a feature amount related to a distribution of pixel values of the image based on the detection target image subjected to the overall normalization process, and using the feature amount, Face image candidate detection means for detecting a face image candidate in the detection target image, and a pixel value on the image for the partial image corresponding to the face image candidate in the detection target image subjected to the overall normalization process Local normalization means for performing local normalization processing for converting the dispersion of pixel values for each local region of a predetermined size in the image region so as to approach a predetermined level, and the portion subjected to the local normalization processing A feature amount related to the distribution of pixel values of the image is calculated based on the image, and the feature amount is used to function as a determination unit that determines whether a face image candidate corresponding to the partial image is a face image. Is for.

  Here, as the “total normalization process”, for example, the detection target image has an sRGB color space, that is, the gamma value (γout) of the image output device is 2.2, and the gamma value at the time of image acquisition is 0. When it is assumed that the image is acquired with a standard of 45 (= 1 / γout), the pixel value in the entire image of the detection target image is multiplied by γout (= 2.2) and then the logarithm is further increased. A process of conversion according to a conversion curve (see FIG. 15) can be considered.

  Note that the “total normalization process” may always perform pixel value conversion according to a fixed conversion curve, but for each detection target image, the distribution of the pixel values of the entire image and the histogram of the pixel values are calculated. The conversion curve for obtaining the image contrast closer to the level suitable for the face detection process is selected or derived by calculation based on the information, and the pixel value is converted according to the conversion curve. Also good.

  The “degree of dispersion of pixel values” means the degree of dispersion of pixel values. For example, in addition to a so-called mathematical dispersion value of pixel values, the difference between the maximum value and the minimum value of pixel values. It can be a value or the like. Further, the “predetermined level” here is one of the values that the dispersion value or the difference value can take, and an image whose contrast that originally represents the face frame, cheek, background sky, etc. is flat. And the difference value or the like corresponding to a boundary that can distinguish the image from other images.

  The “predetermined representative value in the pixel value statistics” is a central value of the distribution representing the characteristics of the distribution of the pixel value. For example, the statistical average value and the median value of the pixel value , Intermediate values, mode values, and the like.

  The “local region” is preferably a rectangular region for convenience, but may be a shape such as a perfect circle or an ellipse.

  In the face detection apparatus of the present invention, the discrimination means is a discriminator in which features relating to the distribution of pixel values of a face image are learned in advance from a plurality of different face sample images, and the local normalization process is performed. The image processing apparatus may include a discriminator that determines whether or not the face image candidate corresponding to the partial image is a face image using the feature amount related to the partial image.

  Further, the local normalization means calculates a degree of dispersion of pixel values for each local area in the partial image, and as the local normalization process, the local area where the calculated degree of dispersion is a predetermined level or more The first luminance gradation conversion process is performed to bring the degree of dispersion of pixel values closer to a certain level higher than the predetermined level, and for the local region where the calculated degree of dispersion is less than the predetermined level Further, a second luminance gradation conversion process may be performed in which the degree of dispersion of pixel values is suppressed to a level lower than the predetermined level.

  Here, as the “local normalization process”, each pixel in the partial image is sequentially set as a target pixel, and for each target pixel, a pixel value in a local region having a predetermined size centered on the target pixel. When the variance is greater than or equal to a threshold value corresponding to the predetermined level, as the first luminance gradation conversion process, as the variance is larger than a reference value corresponding to the certain level, the attention The difference between the pixel value of the pixel and a predetermined statistical value of the pixel value in the local area centered on the pixel of interest is reduced, and the pixel value of the pixel of interest is smaller as the variance is smaller than the reference value. And gradation conversion for increasing the difference between the predetermined representative value and the predetermined representative value.

  Here, the “predetermined level” may be changed according to the whole or a part of luminance in the local region. For example, in the above-described local normalization process in which gradation conversion is performed for each target pixel, the threshold value may be changed according to the pixel value of the target pixel. That is, the threshold corresponding to the predetermined level may be set higher when the luminance of the pixel of interest is relatively high, and may be set lower when the luminance is relatively low. In this way, a face that exists in a low-luminance (dark) region with a low contrast (small dispersion value) can be correctly normalized.

  According to the face detection method and apparatus and the program therefor according to the present invention, the stage for extracting face candidates by performing a relatively rough face detection process on the detection target image, that is, the processing target area is wide. While high speed is important, it is easy to extract the face candidates here, so in a situation where reliability is not so required, it is easily affected by oblique light and background in the image, but on the other hand, the processing time is short In the refinement stage that adopts the overall normalization process that has the advantage of the above, and performs a fine discrimination process on the extracted image near the face candidate to determine whether it is a true face, that is, among the face candidates While the importance of reliability is important because non-faces must be rigorously excluded from the image, the area to be processed is limited, so high speed is not so required. However, the processing time is long, but the local normalization process has the advantage that it is less susceptible to the effects of oblique light and background in the image. Normalization processing having characteristics suitable for satisfying high speed can be applied, and the advantages of each normalization processing can be further utilized to enable highly accurate and efficient face detection.

  Hereinafter, embodiments of the present invention will be described. FIG. 1 is a schematic block diagram showing the configuration of a face detection system to which the face detection apparatus of the present invention is applied. This face detection system detects a face included in a digital image regardless of the position, size, orientation, and rotation direction of the face. As shown in FIG. 1, the face detection system 1 converts a resolution image group S1 (= S1_1) composed of a plurality of images having different resolutions (hereinafter referred to as resolution images) by multi-resolution the input image S0 that is a face detection target. , S1_2,..., S1_n) and a pre-processing for improving the accuracy of the face detection process to be executed later, a contrast suitable for the face detection process is set for each resolution image. A conversion curve that brings the pixel value of the entire image close to a value representing the logarithm of the luminance of the subject in this image for each of the resolution image groups S1 in order to approach the level (a level suitable for extracting the performance of the discriminator described later). To obtain a resolution image group S1 ′ (= S1′_1, S1′_2,..., S1′_n) whose entire normalization has been performed. By performing face detection processing on each of the normalization unit 20 and the overall normalized resolution image group S1 ′, images representing faces included in the resolution images of the resolution image group S1 ′ (hereinafter referred to as face images) ) For each of the face detection unit 30 for detecting S2 and the face image S2 detected on each resolution image, it is determined from the positional relationship whether or not the same face is detected in duplicate. And a duplication detection determination unit 40 that arranges and obtains a face image S3 without duplication detection.

  The multi-resolution conversion unit 10 converts the resolution (image size) of the input image S0 to normalize the resolution to a predetermined resolution, for example, an image having a rectangular size with a short side of 416 pixels. An image S0 ′ is obtained. Then, by further performing resolution conversion based on the standardized input image S0 ′, a plurality of resolution images having different resolutions are generated, and a resolution image group S1 is obtained. The reason why such a resolution image group is generated is that the size of the face included in the input image is usually unknown, while the size of the face to be detected (image size) is determined by a discriminator described later. Since it is fixed to a certain size in relation to the generation method, in order to detect faces of different sizes, each partial image of a predetermined size is cut out while shifting the position on an image with a different resolution. This is because it is necessary to determine whether the face is non-face. Specifically, as shown in FIG. 2, a standardized input image S0 ′ is set as a basic resolution image S1_1, and a resolution image S1_2 having a size that is −1/3 times the size of the resolution image S1_1, A resolution image S1_3 having a size of 2−1 / 3 times the size of the resolution image S1_2 (2−2 / 3 times the size of the basic image S1_1) is generated first, and then the resolution image S1_1, A plurality of resolutions are generated by repeatedly performing a process of generating a resolution image obtained by reducing each of S1_2 and S1_3 to 1/2 times the size, and generating a resolution image obtained by further reducing the reduced resolution image to 1/2 times the size. A predetermined number of images are generated. In this way, the reduction processing of 1/2 times that does not require the interpolation processing of the pixel value representing the luminance is the main processing, and the size is reduced by 2−1 / 3 times from the basic resolution image. A plurality of images can be generated at high speed. For example, when the resolution image S1_1 has a rectangular size of 416 pixels on the short side, the resolution images S1_2, S1_3,... Have 330 pixels, 262 pixels, 208 pixels, 165 pixels, 131 pixels, and 104 pixels on the short sides, respectively. , 82 pixels, 65 pixels,..., And can generate a plurality of resolution images reduced by a factor of 2 to −1/3. Note that an image generated without interpolating pixel values in this way has a strong tendency to retain the characteristics of the original image pattern as it is, and is preferable in that an improvement in accuracy can be expected in face detection processing.

  The overall normalization unit 20 performs overall normalization processing on each of the resolution image groups S1. Specifically, for example, as shown in FIG. 15, pixel values are converted into so-called inverse gamma in the sRGB space. A process of converting pixel values in the entire image can be considered according to a conversion curve (look-up table) that takes a logarithm after conversion (= 2.2). This is due to the following reason.

  The light intensity I observed as an image is usually expressed as the product of the reflectance R of the subject and the intensity L of the light source (I = R × L). Therefore, when the intensity L of the light source changes, the light intensity I observed as an image also changes. However, if only the reflectance R of the subject can be evaluated, it does not depend on the intensity L of the light source. It is possible to perform highly accurate face discrimination that is not affected by the brightness of the image.

Here, when the intensity of the light source is L, the light intensity observed from the portion with the reflectance R1 on the subject is I1, and the light intensity observed from the portion with the reflectance R2 on the subject is I2. In the space where each logarithm is taken, the following equation holds.
log (I1) −log (I2) = log (R1 × L) −log (R2 × L) = log (R1) + log (L) − (log (R2) + log (L)) = log (R1) −log (R2) = log (R1 / R2)

  In other words, logarithmic conversion of pixel values in an image results in conversion into a space where the reflectance ratio is expressed as a difference. In such a space, only the reflectance of the subject that does not depend on the intensity L of the light source is evaluated. It becomes possible to do. In other words, it is possible to align different contrasts (here, the pixel value difference itself) depending on the brightness in the image.

  On the other hand, the color space of an image acquired by a device such as a general digital camera is sRGB. sRGB is an international standard color space that defines and unifies color, saturation, etc., in order to unify the differences in color reproduction between devices. In this color space, the gamma value (γout) is 2. The image pixel value is a value obtained by raising the input luminance to 1 / γout (= 0.45) in order to enable proper color reproduction in the .2 image output device.

  Therefore, the pixel value in the entire image is converted according to a so-called inverse gamma conversion, that is, according to a conversion curve that takes a logarithm after being raised to the power of 2.2, thereby evaluating only by the reflectance of the subject independent of the intensity of the light source. Can be performed properly.

  In other words, such an overall normalization process is a process of converting pixel values in the entire image according to a conversion curve for converting a specific color space into a color space having different characteristics. it can.

  By applying such processing to the detection target image, different contrasts can be provided depending on the brightness in the image, and the accuracy of the face detection processing is improved. The overall normalization process is characterized in that the processing result is easily influenced by the difference in oblique light, background, and input modality in the detection target image, but the processing time is short.

  The face detection unit 30 performs a face detection process on each of the resolution image groups S1 ′ subjected to the overall normalization process by the overall normalization unit 20, and detects a face image S2 in each resolution image. FIG. 3 is a block diagram showing the configuration of the face detection unit 30. As shown in FIG. As shown in FIG. 3, the face detection unit 30 controls each unit described later to mainly perform sequence control in the face detection process, and the resolution used for the face detection process from the resolution image group S1 ′. A resolution image selection unit 32 that sequentially selects images in order from the smallest size, and a sub-window for cutting out a partial image W that is a determination target of whether or not the image is a face image in the resolution image selected by the resolution image selection unit 32 A sub-window setting unit 33 that sequentially sets while shifting the position, a first discriminator group 34 and a second discriminator group 35 that discriminate whether or not the cut out partial image W is a face image, It comprises a local normalization unit 36 that performs a local normalization process on the partial image W input to the second discriminator group 35.

  The detection control unit 31 roughly detects a face image candidate as a face image candidate for each image in the resolution image group S1 ′, and further extracts a true face image S2 from the face image candidates. In order to perform stepwise face detection processing, the resolution image selection unit 32 and the sub window setting unit 33 are controlled. For example, as appropriate, the resolution image selection unit 32 is instructed to select a resolution image, the subwindow setting unit 33 is instructed to set subwindow settings, and the obtained detection result is used as a duplicate detection determination unit 40. Or output to The sub-window setting conditions include the range on the image where the sub-window is set, the sub-window movement interval (detection roughness), and the classifier group used for discrimination (rough / high-precision detection mode). It is.

  Under the control of the detection control unit 31, the resolution image selection unit 32 sequentially selects resolution images to be subjected to face detection processing from the resolution image group S1 ′ in ascending order of size (in order of coarse resolution). Note that the face detection method in the present embodiment determines whether the partial image W is a face image for the partial images W of the same size sequentially cut out on each resolution image, thereby determining the face in the input image S0. Therefore, the resolution image selection unit 32 sets the face size to be detected in the input image S0 while changing the size of the face to be detected from large to small. It can be said that it is equivalent to what is set while changing.

  The sub window setting unit 33 sequentially sets the sub window on the resolution image selected by the resolution image selection unit 32 based on the sub window setting condition set by the detection control unit 31. For example, in the case of performing the rough detection, the sub-window for cutting out the partial image W having a predetermined size, that is, a 32 × 32 pixel size in the selected resolution image is moved by a predetermined number of pixels, for example, 5 pixels. The cut partial image W is input to the first discriminator group 34. As will be described later, each discriminator constituting the discriminator group discriminates a face image of a face in a predetermined direction and a vertical direction, and in this way, in any direction and a vertical direction. It becomes possible to discriminate a face image of a certain face. Further, when narrowing down the face image candidates, the resolution image is limited to a predetermined area including the face image candidates within the predetermined size, and the sub-window is set at a shorter interval, for example, one pixel at a time. The image is sequentially set while being moved, the partial image W is cut out in the same manner as described above, and the cut out partial image W is input to the second discriminator group 35 via the local normalization unit 36.

  The first discriminator group 34 is a discriminator group that discriminates whether or not the partial image W is a face image at a relatively high speed, and is used to roughly detect face image candidates in the resolution image. As shown in FIG. 4, the first discriminator group 34 includes a plurality of types of discriminator groups having different face orientations, that is, a first front face discriminator group that mainly discriminates front faces. 34_F, a first left profile classifier group 34_L that mainly determines the left profile and a first right profile classifier group 34_R that mainly determines the right profile are connected in parallel. Further, each of these three types of classifier groups is a classifier corresponding to a total of 12 directions in which the top and bottom directions of the distinguishable face differ by 30 degrees with respect to the top and bottom direction of the image, that is, the first front face classifier group 34_F is a discriminator 34_F30, 34_F60,..., 34_F330, the first left side face discriminator group 34_L is a discriminator 34_L30, 34_L60,. 34_R30, 34_R60,..., 34_R330.

  On the other hand, the second discriminator group 35 determines whether or not the partial image W (strictly, the partial image W ′ subjected to the local normalization processing by the local normalization unit 36) is a face image. And is used to extract (narrow down) the true face image S2 from the face image candidates by performing finer detection processing on the face image candidates detected by the rough detection described above. . Similar to the first discriminator group, the second discriminator group 35 discriminates a plurality of types of discriminator groups having different face directions that can be discriminated as shown in FIG. The second front face classifier group 35_F, the second left side face classifier group 35_L that mainly discriminates the left side face, and the second right side face classifier group 35_R that mainly discriminates the right side face are connected in parallel. It is a configuration. Furthermore, these three types of classifier groups, like the first classifier group, are classifiers each corresponding to a total of 12 directions in which the top and bottom directions of the distinguishable faces differ by 30 degrees with respect to the top and bottom direction of the image. The second front face discriminator group 35_F includes discriminators 35_F30, 35_F60,..., 35_F330, and the second left side face discriminator group 35_L includes discriminators 35_L30, 35_L60,. The right side face classifier group 35_R includes classifiers 35_R30, 35_R60,..., 35_R330.

  When the partial normalization unit 36 inputs the partial image W corresponding to the face image candidate to the second discriminator group in order to narrow down the extracted face image candidates, the local normalization unit 36 performs pre-processing on the partial image W. Thus, local normalization processing is performed to suppress contrast variation in a local region on the image. In other words, the local normalization unit 36 applies, for each local region in the partial image W, the local region in which the degree of dispersion of pixel values representing luminance is equal to or higher than a predetermined level. A first luminance gradation conversion process is performed to bring the degree of dispersion closer to a certain level higher than the predetermined level, and the degree of dispersion is set for a local region where the degree of dispersion of pixel values is less than the predetermined level. A second luminance gradation conversion process is performed to suppress the level to a level lower than the predetermined level. This local normalization process has a long processing time, but has a feature that the influence on the determination result due to the difference in oblique light, background, and input modality in the detection target image is small. Here, specific processing in the local normalization unit 36 will be described.

FIG. 12 is a diagram showing the concept of local normalization processing, and FIG. 13 is a diagram showing a processing flow in the local normalization unit 36. Expressions (1) and (2) are gradation conversion expressions for pixel values for the local normalization process.

Here, X is the pixel value of the pixel of interest, X ′ is the pixel value after conversion of the pixel of interest, mlocal is the average of the pixel values in the local region centered on the pixel of interest, Vlocal is the variance of the pixel values in this local region, SDlocal is a standard deviation of pixel values in this local area, (C1 × C1) is a reference value corresponding to the above-mentioned constant level, C2 is a threshold value corresponding to the above-mentioned predetermined level, and SDc is a predetermined constant. In the present embodiment, the number of gradations of luminance is 8 bits, and the possible pixel values are 0 to 255.

  As shown in FIG. 13, the local normalization unit 36 sets one pixel in the partial image W as a target pixel (step S31), and has a predetermined size centered on the target pixel, for example, an 11 × 11 pixel size. A variance Vlocal of pixel values in the local region is calculated (step S32), and it is determined whether or not the variance Vlocal is equal to or greater than a threshold C2 corresponding to the predetermined level (step S33). If it is determined in step S33 that the variance Vlocal is greater than or equal to the threshold value C2, the variance Vlocal is larger than the reference value (C1 × C1) corresponding to the certain level as the first luminance gradation conversion process. The tone conversion that decreases the difference between the pixel value X of the target pixel and the average mlocal, and increases the difference between the pixel value X of the target pixel and the average mlocal as the variance mlocal is smaller than the reference value (C1 × C1). Is performed according to the equation (1) (step S34). On the other hand, if it is determined in step S33 that the variance Vlocal is less than the threshold value C2, linear tone conversion that does not depend on the variance Vlocal is performed according to equation (2) as the second luminance tone conversion processing. This is performed (step S35). Then, it is determined whether or not the target pixel set in step S31 is the last pixel (step S36). If it is determined in step S36 that the target pixel is not the last pixel, the process returns to step S31, and the next pixel on the same partial image is set as the target pixel. On the other hand, if it is determined in step S36 that the target pixel is the last pixel, the local normalization for the partial image is terminated. As described above, by repeating the processes of steps S31 to S36, a partial image W ′ obtained by local normalization of the entire partial image is obtained.

  Note that the predetermined level may be changed according to the whole or a part of luminance in the local region. For example, in the normalization process in which gradation conversion is performed for each target pixel, the threshold value C2 may be changed according to the pixel value of the target pixel. That is, the threshold value C2 corresponding to the predetermined level may be set higher when the luminance of the target pixel is relatively high, and may be set lower when the luminance is relatively low. In this way, it is possible to correctly normalize a face that exists in a low-brightness, so-called dark area with low contrast (a state in which the dispersion of pixel values is small).

  Each discriminator has a cascade structure in which a plurality of weak discriminators WC are linearly coupled as shown in FIG. 4, and the weak discriminator has a pixel value (luminance) of the partial image W. At least one feature amount related to the distribution is calculated, and using this feature amount, it is determined whether or not the partial image W is a face image.

  The first and second discriminator groups 34 and 35 both have three types of main face directions that can be discriminated: a front face, a left side face, and a right side face. In order to increase the accuracy, a discriminator for discriminating each of the right oblique face and the left oblique face may be further provided.

  Here, a specific process in each discriminator will be described. FIG. 5 shows a general processing flow in each discriminator, and FIG. 6 shows a processing flow by each weak discriminator therein.

  First, the first weak discriminator WC discriminates whether or not the partial image W is a face with respect to the partial image W of a predetermined size cut out on the predetermined resolution image S1′_i (step SS1). . Specifically, as shown in FIG. 7, the first weak discriminator WC applies a partial image W of a predetermined size cut out on the resolution image S1′_i, that is, an image of 32 × 32 pixel size. 4-neighbor pixel average (processing that divides an image into a plurality of blocks for each 2 × 2 pixel size and sets an average value of pixel values of four pixels of each block as a pixel value of one pixel corresponding to the block) By doing this, an image with a size of 16 × 16 pixels and an image with a reduced size of 8 × 8 pixels are obtained, and a predetermined two points set in the plane of these three images are taken as one pair, and a plurality of types of pairs are used. A difference value of pixel values (luminance) between two points in each pair constituting one pair group is calculated, and a combination of these difference values is used as a feature amount (step SS1-1). The predetermined two points of each pair are, for example, two predetermined points arranged in the vertical direction and two predetermined points arranged in the horizontal direction so as to reflect the characteristics of the facial shading on the image. Then, a score is calculated by referring to a predetermined score table according to a combination of difference values as feature amounts (step SS1-2), and the score calculated by itself is added to the score calculated by the previous weak discriminator. The accumulated score is calculated (step SS1-3). However, since the first weak discriminator WC1 has no previous weak discriminator, the score calculated by itself is used as the cumulative score. It is determined whether or not the partial image is a face depending on whether or not the accumulated score is equal to or greater than a predetermined threshold (step SS1-4). Here, when the partial image W is determined to be a face, the process proceeds to determination by the next weak classifier WC2 (step SS2). When the partial image W is determined to be a non-face, the partial image is immediately non- The face is determined (step SSB), and the process ends.

  Also in step SS2, as in step SS1, the second weak classifier WC calculates the above-described feature amount representing the feature on the image based on the partial image (step SS2-1), and refers to the score table. Then, a score is calculated from the feature amount (step SS2-2). Then, the score calculated by itself is added to the cumulative score calculated by the immediately preceding first weak discriminator WC to update the cumulative score (step SS2-3), and whether or not the cumulative score is equal to or greater than a predetermined threshold value. To determine whether the partial image W is a face (step SS2-4). Again, when the partial image W is determined to be a face, the process proceeds to determination by the next third weak classifier WC (step SS3). When the partial image W is determined to be a non-face, the partial image W is Immediately, a non-face is determined (step SSB), and the process ends. Thus, when the partial image W is determined to be a face in all N weak classifiers WC constituting the classifier, the partial image W is finally extracted as a face image candidate (step SSA). .

  Each of the discriminators is a discriminator including a plurality of weak discriminators WC defined by unique feature type, score table, and threshold value, and discriminates faces in a predetermined direction and a vertical direction, respectively. .

  The duplicate detection determination unit 40 represents the same face among the face images detected on each resolution image of the resolution image group S1 ′ based on the position information of the true face image S2 detected by the face detection unit 30. A process of combining images, that is, face images detected redundantly, as one face image is performed, and a true face image S3 detected in the input image S0 is output. Depending on the learning method, the discriminator generally has a certain range in the size of the face that can be detected with respect to the size of the partial image W. Therefore, the same face in a plurality of resolution images having adjacent resolution levels. This is because there are cases where images representing the same are detected in duplicate.

  In the present embodiment, the overall normalization unit 20 functions as the overall normalization unit of the present invention, and the local normalization unit 36 functions as the local normalization unit of the present invention. The detection control unit 31, the resolution image selection unit, 32, the sub-window setting unit 33 and the first discriminator group 34 function as face image candidate detecting means of the present invention, and the detection control unit 31, the sub-window setting unit 33 and the second discriminator group 35 are discriminating means of the present invention. Function as.

  Next, the flow of processing in the face detection system 1 will be described. FIG. 9 is a flowchart showing the flow of processing in the face detection system. As shown in FIG. 9, when the input image S0 is supplied to the multi-resolution converting unit 10 (step S1), an image S0 ′ in which the image size of the input image S0 is converted to a predetermined size is generated. A resolution image group S1 composed of a plurality of resolution images reduced in size (resolution) by 2 to 1/3 times from S0 'is generated (step S2). Then, the overall normalization unit 20 performs overall normalization processing for suppressing the variation in contrast of the entire image for each resolution image group S1, that is, the pixel value of the entire image represents the logarithm of the luminance of the subject in the image. A conversion process is performed according to a conversion curve that approximates the value, and the overall normalized resolution image group S1 ′ is obtained (step S3). In the face detection unit 30, the resolution image selection unit 32 that has received an instruction from the detection control unit 31 starts from the resolution image group S 1 ′ in ascending order of image size, that is, S 1 ′ _n, S 1 ′ _n−1, ..., a predetermined resolution image S1'_i is selected in the order of S1'_1 (step S4). Next, the detection control unit 31 sets sub-window setting conditions for setting the detection mode to a rough detection mode for the sub-window setting unit 33. As a result, the sub-window setting unit 33 sets the sub-window on the resolution image S1′_i while moving the sub-window at a wider pitch, for example, at an interval of 5 pixels, and sequentially cuts out the partial images W of a predetermined size (step S5). W is input to the first discriminator group 34 (step S6). The first discriminator group 34 discriminates the sequentially input partial images W using the 36 kinds of discriminators described above, and the detection control unit 31 acquires the discrimination result R (step S7). . Then, the detection control unit 31 determines whether or not the currently cut out partial image W is a partial image positioned in the last order (step S8), and determines that the partial image W is the last partial image. If it is determined that the partial image W is not the last partial image, the process returns to step S5 to cut out a new partial image W. In this way, face image candidates for the resolution image S1′_i are roughly detected.

  When the rough detection of the face image candidate is completed, the detection control unit 31 determines whether or not the face image candidate is detected. If it is determined that the face image candidate is detected, the narrowing mode is further reduced. If it is determined that no face image candidate has been detected, detection for the currently selected resolution image S1′_i is performed without performing detection in the narrow-down mode. Ends, and the process proceeds to step S14.

  In step S10, the detection control unit 31 sets a sub-window setting condition in which the detection target region is limited to a predetermined size region including the face image candidate, and the detection mode is set as a narrowing mode, with respect to the sub-window setting unit 33. To do. Accordingly, the sub window setting unit 33 sets the sub window in the vicinity of the face image candidate while moving the sub window with a narrow pitch, for example, one pixel at a time, and sequentially cuts out the partial images W of a predetermined size (step S10). Input to the normalization unit 36. The local normalization unit 36 performs luminance gradation conversion that brings the variance close to a certain level for an area where the variance of the pixel values of the partial image W is equal to or greater than a predetermined threshold, and the variance of the pixel values is the predetermined threshold. For the region below the threshold level, local normalization is performed to perform luminance gradation conversion that suppresses the variance to a level lower than the predetermined level (step S11), and the locally normalized partial image W ′ is converted into the second discriminator. Input to the group 35 (step S12). For the partial images W ′ that are sequentially input, the second discriminator group 35 includes three types of face orientations: front face, right profile, and left profile, and 12 types that differ by 30 degrees in the vertical direction. The 36 types of faces are discriminated using each discriminator, and the detection control unit 31 acquires the discrimination result R (step S13). Then, the detection control unit 31 determines whether or not the currently extracted partial image W is a partial image positioned in the last order (step S14), and determines that the partial image W is the last partial image. If it is determined that the partial image W is not the last partial image, the process returns to step S10, and a new partial image W is cut out. In this way, the detected face image candidates are narrowed down, and the true face image S2 in the resolution image S1′_i is extracted.

  When the detection of the narrow-down mode in the vicinity region of the face image candidate is completed, the detection control unit 31 determines whether or not the currently selected resolution image S1′_i is an image positioned in the last order (step S15). ), If it is determined that the image is the last resolution image, the detection process is terminated, and the process proceeds to duplicate detection determination (step S16). On the other hand, when it is determined that the resolution image is not the last resolution image, the process returns to step S10, and the resolution image selection unit 32 sets the resolution image S1′_i− that is one step larger than the currently selected resolution image S1′_i. 1 is selected, and face image detection is further performed.

  In this way, the face image S2 in each resolution image can be detected by repeating the processing from step S4 to S15. FIG. 8 is a diagram showing how face detection is performed by selecting resolution images in ascending order of size.

  In step S16, the overlap detection determination unit 40 performs a process of combining each face image detected in the true face image S2 as one face image, and the true face image S3 detected in the input image S0. Is output.

  Next, a learning method (generation method) of the discriminator will be described. Note that learning is performed for each type of classifier, that is, for each combination of the orientation of the face to be determined and the vertical direction.

  The sample image group to be learned is a plurality of sample images (face sample image group) that are known to be faces and are not faces, standardized at a predetermined size, for example, 32 × 32 pixel size. It consists of a plurality of known sample images (non-face sample image group). As a sample image that is known to be a face, an image in which the face orientation is the same as the face orientation to be discriminated by the discriminator and the face orientations are aligned is used. A sample image that is known to be a face is obtained by scaling in steps of 0.1 times in the range of 0.7 to 1.2 times in length and / or width for each sample image. For each sample image to be obtained, a plurality of deformation variations obtained by rotating stepwise in units of 3 degrees within a range of ± 15 degrees on a plane is used. At this time, the face sample image is standardized in size and position so that the eye position is at a predetermined position, and the above-described rotation and scaling on the plane are performed based on the eye position. For example, in the case of a d × d size front face sample image, as shown in FIG. 14, the positions of both eyes are 1/4 d inward from the upper left vertex and the upper right vertex of the sample image, respectively. The face size and position are normalized so as to come to each position moved 1 / 4d downward, and the above-mentioned rotation and scaling on the plane are performed around the middle point of both eyes.

  Such face sample image groups are prepared for a total of 36 types of 12 types, each of which has a top-to-bottom direction of 30 degrees for each of the front face, right profile, and left profile. Each of these 36 types of face sample image groups and the non-face sample image group is used to learn a classifier for each type to generate 36 types of classifiers. The specific learning method will be described below.

  FIG. 10 is a flowchart showing a learning method of the classifier. It is assumed that each sample image constituting the face sample image group and the non-face sample image group has been subjected in advance to data conversion processing equivalent to the data conversion processing by the data conversion processing unit 20 described above.

  Each of these sample images is assigned a weight or importance. First, the initial value of the weight of all sample images is set equal to 1 (step S21).

  Next, when a plurality of pairs of groups consisting of a plurality of pairs are set with a predetermined two points set in the plane of the sample image and the reduced image as one pair, each of the plurality of types of pairs is weak. A separate device is created (step S22). Here, each weak discriminator sets one pair group consisting of a plurality of pairs with a predetermined two points set in the plane of the partial image cut out in the sub-window W and the reduced image as one pair. This provides a reference for discriminating between a face image and a non-face image using a combination of difference values of pixel values (luminance) between two points in each pair constituting this one pair group. . In the present embodiment, a histogram for a combination of pixel value difference values between two points in each pair constituting one pair group is used as the basis of the score table of the weak classifier.

  The creation of a classifier will be described with reference to FIG. As shown in the sample image on the left side of FIG. 11, two points of each pair constituting the pair group for creating this discriminator are a plurality of sample images that are known to be faces. The right eye on the reduced image of 16 × 16 pixel size in which the point in the center of the right eye is P1, the point in the right cheek part is P2, the point in the part between the eyebrows is P3, and the sample image is reduced by an average of four neighboring pixels The point at the center of P4, the point at the cheek on the right side is P5, and the point at the forehead part on the reduced image of 8 × 8 pixel size reduced by the average of 4 neighboring pixels is P6, the mouth part A certain point is P7, and there are five pairs of P1-P2, P1-P3, P4-P5, P4-P6, and P6-P7. Note that the coordinate positions of the two points of each pair constituting one pair group for creating a certain classifier are the same in all sample images. For all sample images that are known to be faces, combinations of pixel value difference values between two points of each of the five pairs are obtained, and a histogram thereof is created. Here, the value that can be taken as a combination of the difference values of the pixel values depends on the number of luminance gradations of the image, but if it is a 16-bit gradation, there are 65536 kinds of difference values of one pixel value, As a whole, the number of gradations is (the number of pairs), that is, 65536 to the fifth power, and a large number of samples, time, and memory are required for learning and detection. For this reason, in the present embodiment, the difference value of the pixel value is divided by an appropriate numerical value width and quantized to be n-valued (eg, n = 100).

  As a result, the number of combinations of difference values of pixel values is n to the fifth power, so that the number of data representing combinations of difference values of pixel values can be reduced.

  Similarly, histograms are created for a plurality of sample images that are known not to be faces. For sample images that are known not to be faces, positions corresponding to the positions of the two predetermined points of each pair on the sample image that is known to be a face (similarly, reference numerals P1 to P7 are used). ) Is used. A histogram obtained by taking the logarithm of the ratio of the frequency values indicated by these two histograms and representing the histogram is the histogram used as the basis of the score table of the weak discriminator shown on the rightmost side of FIG. The value of each vertical axis indicated by the histogram of the weak classifier is hereinafter referred to as a discrimination point. According to this weak discriminator, an image showing the distribution of combinations of pixel value difference values corresponding to positive discrimination points is highly likely to be a face, and the possibility increases as the absolute value of the discrimination point increases. It can be said. Conversely, an image showing a distribution of combinations of difference values of pixel values corresponding to negative discrimination points is highly likely not to be a face, and the possibility increases as the absolute value of the discrimination point increases. In step S22, a plurality of weak discriminators in the above-described histogram format are created for combinations of pixel value difference values between predetermined two points of each pair constituting a plurality of types of pair groups that can be used for discrimination.

  Subsequently, the most effective weak discriminator for discriminating whether or not the image is a face is selected from the plurality of weak semi-divided devices created in step S22. The most effective weak classifier is selected in consideration of the weight of each sample image. In this example, the weighted correct answer rates of the weak classifiers are compared, and the weak classifier showing the highest weighted correct answer rate is selected (step S23). That is, in the first step S23, since the weight of each sample image is equal to 1, the one with the largest number of sample images for which it is simply determined correctly whether or not the image is a face by the weak classifier is as follows: Selected as the most effective weak classifier. On the other hand, in the second step S23 after the weight of each sample image is updated in step S25, which will be described later, a sample image with a weight of 1, a sample image with a weight greater than 1, and a sample image with a weight less than 1 The sample images having a weight greater than 1 are counted more in the evaluation of the correct answer rate because the weight is larger than the sample images having a weight of 1. Thereby, in step S23 after the second time, more emphasis is placed on correctly determining a sample image having a large weight than a sample image having a small weight.

  Next, the correct answer rate of the combination of weak classifiers selected so far, that is, using the weak classifiers selected so far in combination (in the learning stage, the weak classifiers do not necessarily need to be linearly combined. ) It is ascertained whether the result of determining whether or not each sample image is a face image has exceeded a predetermined threshold value at a rate that matches the answer of whether or not it is actually a face image (step) S24). Here, the current weighted sample image group or the sample image group with equal weight may be used for evaluating the correct answer rate of the combination of weak classifiers. When the predetermined threshold value is exceeded, learning is terminated because it is possible to determine whether the image is a face with a sufficiently high probability by using the weak classifier selected so far. If it is equal to or less than the predetermined threshold value, the process proceeds to step S26 in order to select an additional weak classifier to be used in combination with the weak classifier selected so far.

  In step S26, the weak classifier selected in the most recent step S23 is excluded so as not to be selected again.

  Next, the weight of the sample image in which the weak discriminator selected in the most recent step S23 could not correctly discriminate whether it is a face is increased, and the sample image in which whether the image is a face can be discriminated correctly is increased. Is reduced (step S25). The reason for increasing or decreasing the weight in this way is that in the selection of the next weak classifier, importance is placed on images that could not be correctly determined by the already selected weak classifier, and whether or not those images are faces is correct. This is because a weak discriminator that can be discriminated is selected to enhance the effect of the combination of the weak discriminators.

  Subsequently, the process returns to step S23, and the next effective weak classifier is selected based on the weighted correct answer rate as described above.

  As a weak discriminator suitable for discriminating whether or not a face is repeated by repeating the above steps S23 to S26, the difference value of the pixel value between two predetermined points of each pair constituting a specific pair group If the weak discriminator corresponding to the combination is selected and the correct answer rate confirmed in step S24 exceeds the threshold value, the type of the weak discriminator used for discriminating whether the face is a face and the discrimination condition are determined. (Step S27), thereby completing the learning. The selected weak classifiers are linearly combined in descending order of the weighted correct answer rate to constitute one classifier. For each weak classifier, a score table for calculating a score according to a combination of pixel value difference values is generated based on the obtained histogram. Note that the histogram itself can also be used as a score table. In this case, the discrimination point of the histogram is directly used as a score.

  In this way, by performing learning for each face sample image group, the 36 types of discriminators described above are generated.

  In the case of adopting the above learning method, the weak classifier uses a combination of difference values of pixel values between two predetermined points of each pair constituting a specific pair group, and a face image and a non-face image. Is not limited to the above-described histogram format, and may be anything, for example, binary data, a threshold value, a function, or the like. Further, even in the same histogram format, a histogram or the like indicating the distribution of difference values between the two histograms shown in the center of FIG. 11 may be used.

  Further, the learning method is not limited to the above method, and other machine learning methods such as a neural network can be used.

  As described above, according to the face detection system according to the embodiment of the present invention, a stage in which a relatively rough face detection process is performed on a detection target image to extract face candidates, that is, a region to be processed is included. While high speed is important because it is wide, it is easy to be affected by the oblique light in the image and the background in a scene where reliability is not so required because it is sufficient to extract face candidates here, but on the other hand, processing time Is a narrowing-down stage that adopts an overall normalization process that has the advantage of being short and performs a detailed discrimination process on the extracted image near the face candidate to determine whether it is a true face, that is, a face candidate In the scene where high speed is not required so much because reliability is important because it is necessary to strictly eliminate non-faces from Because it uses a local normalization process that has the advantage of being long in processing time but less susceptible to the effects of oblique light and background in the image, the required process reliability and high speed at each processing stage Normalization processing having characteristics suitable for satisfying the above can be applied, and the advantages of each normalization processing can be further utilized to enable highly accurate and efficient face detection.

  Although the face detection system according to the embodiment of the present invention has been described above, a program for causing a computer to execute each process in a portion corresponding to the face detection device of the present invention in the face detection system is also included in the present invention. This is one of the embodiments. A computer-readable recording medium that records such a program is also one embodiment of the present invention.

Block diagram showing the configuration of the face detection system 1 The figure which shows the process of multiresolution of a detection target image Block diagram showing the configuration of the face detection unit 30 The block diagram which shows the structure of the 1st and 2nd discriminator group. Diagram showing the global processing flow in the classifier The figure which shows the processing flow in the weak classifier The figure for demonstrating calculation of the feature-value in a weak discriminator The figure for demonstrating the change of the resolution image used as a face detection object, and the movement of the subwindow on the image The flowchart which shows the process performed in the face detection system 1 Flow chart showing the learning method of the classifier The figure which shows the method of deriving the histogram of the weak classifier Diagram showing the concept of local normalization processing The figure which shows the processing flow in the local normalization part The figure which shows the sample image of the face standardized so that the position of eyes may be in a predetermined position The figure which shows an example of the conversion curve used for a whole normalization process

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 Face detection system 10 Multi-resolution part 20 Whole normalization part (whole normalization means)
30 face detection unit 31 detection control unit (component of face image candidate detection means / discrimination means)
32 resolution image selection unit (component of face image candidate detection means)
33 Sub-window setting section (components of face image candidate detection means / discrimination means)
34 First classifier group (component of face image candidate detection means)
35 Second discriminator group (component of discriminating means)
36 Local normalization unit (local normalization means)
40 Duplicate detection judgment part

Claims (4)

An overall normalization process that applies an overall normalization process to an input detection target image that is a target for detecting a face according to a conversion curve that approximates the pixel value of the entire image to a value that represents the logarithm of the luminance of the subject in the image Step,
A face image candidate that calculates a feature amount related to a distribution of pixel values of the image based on the detection target image subjected to the overall normalization process, and detects a face image candidate in the detection target image using the feature amount A detection step;
For the partial image corresponding to the face image candidate in the detection target image subjected to the overall normalization process, the pixel value on the image is the degree of dispersion of the pixel value for each local area of a predetermined size in the image area A local normalization step for performing a local normalization process for converting so as to approach a predetermined level;
Based on the partial image subjected to the local normalization process, a feature amount related to the distribution of pixel values of the image is calculated, and using the feature amount, a face image candidate corresponding to the partial image is a face image And a discrimination step for discriminating whether or not. An overall normalization process that applies an overall normalization process to an input detection target image that is a target for detecting a face according to a conversion curve that approximates the pixel value of the entire image to a value that represents the logarithm of the luminance of the subject in the image And
A face image candidate that calculates a feature amount related to a distribution of pixel values of the image based on the detection target image subjected to the overall normalization process, and detects a face image candidate in the detection target image using the feature amount Detection means;
For the partial image corresponding to the face image candidate in the detection target image subjected to the overall normalization process, the pixel value on the image is the degree of dispersion of the pixel value for each local area of a predetermined size in the image area Local normalization means for performing a local normalization process for converting so as to approach a predetermined level;
Based on the partial image subjected to the local normalization process, a feature amount related to the distribution of pixel values of the image is calculated, and using the feature amount, a face image candidate corresponding to the partial image is a face image A face detection apparatus comprising: a determination means for determining whether or not. The discrimination means is
A discriminator that learns in advance a feature related to a distribution of pixel values of a face image from a plurality of different face sample images, using the feature amount related to the partial image subjected to the local normalization process. 3. The face detection apparatus according to claim 2, further comprising a discriminator for discriminating whether or not the face image candidate corresponding to the image is a face image. Computer
An overall normalization process that applies an overall normalization process to an input detection target image that is a target for detecting a face according to a conversion curve that approximates the pixel value of the entire image to a value that represents the logarithm of the luminance of the subject in the image And
A face image candidate that calculates a feature amount related to a distribution of pixel values of the image based on the detection target image subjected to the overall normalization process, and detects a face image candidate in the detection target image using the feature amount Detection means;
With respect to the partial image corresponding to the face image candidate in the detection target image subjected to the overall normalization process, the pixel value on the image is determined to have a predetermined pixel value dispersion for each local region of a predetermined size in the image region. Local normalization means for performing a local normalization process for converting the level closer to the level,
Based on the partial image subjected to the local normalization process, a feature amount related to the distribution of pixel values of the image is calculated, and using the feature amount, a face image candidate corresponding to the partial image is a face image A program for functioning as a determination means for determining whether or not.

Images