KR100338807B1 - Method and apparatus for face detection using classified face images and net type search area - Google Patents

Abstract

The present invention discloses a method and apparatus for face detection using feature point matching of grouped face images and a network structure. According to the present invention, a face detection method of searching for a face of a person includes: (a) generating a model image database by grouping face images to be referred to for finding a face, and (b) size of a face to be found. And detecting a face matching the model image database from the input image using a corresponding search region in the form of a mesh.

Description

[Method and apparatus for face detection using classified face images and net type search area}

The present invention relates to face detection using an image filter, and more particularly, to a face detection method and apparatus using grouped face images and a network-type search area.

The history and results of related researches that detect faces using image filter response values at specific locations are still insignificant at present, and the commercialization and patent application status of these applications are also just beginning. Face finding technology is one of the most fundamental technologies for commercializing or commercializing HCI (Human and Computer Interaction) -related technologies including security systems such as intruder detection and door opening and closing. However, finding a human face quickly and accurately with a small amount of computation in a complex background has a difficulty in solving the mutually opposite problems. Recently, many studies have been actively conducted to solve these problems, and various methods have been proposed.

A typical technique for finding a face is methods using characteristic information of an area or a shape. Method of using facial contours (Fuji Film Corporation US5629752 (1997, 5)), direct comparison of facial areas using neural networks (Siemens EP0474307 (1992, 3)), and skin color, hair color, and facial components. The method using a positional characteristic (Sony Corporation US5812193 (1998, 9), Sanyo Corporation JP8287216 (1996, 1)) etc. are mentioned, for example. However, these methods have severe restrictions on the use environment, such as lighting conditions and backgrounds, have a lot of limitations on application targets, and most of them show insufficient performance for successful commercialization.

An alternative solution to this problem is the PCA method (M.I.T. US5710833 (1998, 1)), which is one of the best performing methods at present. This method extracts only the important information from the face template in the form of one-dimensional information, which is a kind of image compression method. However, this method has to train the face images well aligned, and also has a disadvantage in that the performance changes with a large amount of calculation according to the rotation, size, and facial expression of the face.

Currently, the method that is evaluated as the best is using image filter response values at specific location points (fiducial points) in the face, which is robust to facial expression changes and rotations, and has a relatively flexible advantage in the use environment. However, as a conventional face detection method using a response value of an image filter, a Bunch Graph Matching method ('Automatic digital image recognition system' ZENTRUM FUER NEURO INFORMATIK G, 1995, 6, 29, DE4406020, 'Face Recognition by Elastic Bunch Graph Matching' Internal Report 96-08 ZENTRUM FUER NEURO INFORMATIK G, 1996 In order to find these specific position points well, the comparison operation is performed with all the Jet models in the bunch, which requires a large amount of computation. In addition, as described above, instead of constructing and comparing the bunches needed to find a specific location point in the face, the method simply uses the overall average value of the image filter ('The Role of Topographical Constraints in Face Recognition'). Role) 'Pattern Recognition Letters, 20 (1): 89-96), can benefit from the calculation, but there is a problem in the reliability of face detection.

An object of the present invention is to provide a face detection method which enables reliable face detection while reducing the size of a database to be maintained by using grouped face images and a network type search area.

Another object of the present invention is to provide a face detection apparatus for performing the method.

1 is a block diagram of a face detection apparatus according to the present invention.

2A is a flowchart for explaining a face detection method according to the present invention.

FIG. 2B is a flowchart for describing a method of configuring the model image database referenced in FIG. 2A.

3 shows an example of face image grouping.

4 is a view for explaining a method of calculating an average filter response value using a Gabor filter.

5 shows an example of training a model image for feature point matching using a network structure.

6 (a) to 6 (c) show matching examples for face detection.

In order to achieve the above object, the face detection method according to the present invention to find the face of the person in the input image,

(a) generating a model image database by grouping facial images to be referred to for finding a face; and (b) model image database from an input image using a network-type search region corresponding to a face size to be found. And detecting a face matched with.

In order to achieve the above another object, the face detection device according to the present invention to find a face of a person in the input image,

A face image reference that generates a model image database by grouping face images to be referred to in finding a face, and a mesh-shaped search area corresponding to the size of the face to be searched are matched to the model image database in the input image. And a face detection unit for detecting a face.

Hereinafter, with reference to the accompanying drawings, a face detection method and apparatus according to the present invention will be described.

The present invention is divided into two parts. The training part for creating a model image database by grouping the face images, that is, the model images by features, to be referred to for finding a face, and the input image and the trained model image using a search area in the form of a mesh corresponding to the face size to be searched for. It consists of the part that detects the face in the face image through matching with the database.

1 is a block diagram of a face detection apparatus according to the present invention, which includes a face image reference unit including a grouping unit 102, a first image filter 104, and a model image database 106, and a second image filter 112. ), A face detector including a first searcher 120 and a second searcher 130.

FIG. 2A is a flowchart for explaining a face detection method according to the present invention, and FIG. 2B is a flowchart for explaining a method for configuring a model image database referenced in FIG. 2A.

Referring to FIGS. 2A and 2B, the operation of the apparatus illustrated in FIG. 1 will be described. In operation 208, the grouping unit 102 inputs a model image 100 and groups the features according to features (step 208).

The model images are classified into groups by features in consideration of elements that can characterize the face image, such as a background, a physical element that distorts the face, and an illumination. Unlike face recognition, detecting a face in a background image requires training a model image database in such a way as to distinguish between a face and a non-face part rather than a face-to-face discrimination. Therefore, by grouping face images to be referred to as features, if a face of a person distinguished from the background when projecting a feature domain is viewed as a feature region, the complexity of the region and the breadth of the change are different. It becomes possible to cope.

A network structure corresponding to the size of the face desired to be searched by the user, that is, a search area having a mesh shape, is selected (steps 210 and 200). That is, the feature points are selected to constitute the network structure. The network structure is a structure in which feature points in a face have topographical constraints on each other. The first image filter 104 filters the grouped face images using the selected search region having a mesh shape (operation 212). For each feature point in the network structure, a filter response value for the face images in each group is obtained, and the average filter response value is output as a representative value for the feature point.

The mean filter response of the grouped face images refers to the center of the face groups in the projected feature space. Therefore, it is possible to easily determine whether or not the face with only the distance from the center and the threshold value, compared with all the models one by one compared to the conventional method for determining whether or not the face is less. It also avoids the possibility of errors due to particular exceptional models that may be present when compared to all models. The average filter response of the face image group with this advantage is trained at various resolutions to cope with various face size changes.

The model image database 106 stores face images representing each group obtained by the grouping unit 102 and average filter response values of the grouped face images obtained by the first image filter 104 (see FIG. Step 214). The model image database 106 also stores average filter response values of grouped face images trained at different resolutions to respond to various face size variations. The model image database to be generated is then referred to for face detection and only needs to be retrained or regenerated if necessary.

In the face detector, the second image filter 112 filters the input image 110 (operation 202). The filter response is calculated over the entire input image. The first search unit 120 performs matching between the input image 110 and the model image database 106 using a mesh structure (step 204). Here, the network structure is a search region used to obtain average filter response values of the face images grouped by the first image filter 104.

The first search unit 120 scans the network structure over the entire input image 110, and corresponds to the filter response value of each point in the network structure and the grouped face images included in the model image database 106. Compare the average filter response of the feature points. As a result of the comparison, a face matching the model image database is detected from the input image (step 206). In addition, the first searcher 120 may change the size of the input image and / or the size of the face image in the model image database, and then perform the above-described matching process. In addition, the mesh structure is adjustable in variable size.

Meanwhile, the second search unit 130 may additionally be included. The faces detected by the first searcher 120 are used as face candidate areas, and the second searcher 130 detects faces more precisely. That is, the second search unit 130 extracts an accurate feature point that can identify the face in the face candidate region.

Now, a face detection method according to the present invention and a face detection experiment to which the apparatus is applied will be described.

3 shows an example of face image grouping.

In the experiment for verifying the performance of the present invention, as shown in FIG. 3, the model images were divided into five groups of simple background, backlight, daylight, glasses wearing, and distortion of the eye region image. In addition, the average of the overall image was divided into a total of six groups, and the average filter response values of each group were trained at two scales.

4 is a view for explaining a method of calculating an average filter response value using a Gabor filter.

In the experiment, a Gabor filter with a Gabor wavelet was used. Referring to FIG. 4, a network structure 402, which is a search region, is located in an area including a feature representing a face in a model image. For each point in the network structure, filter response values 420, 422, ... 424 for face images 400, 402, ..., 404 in each group are obtained through Gabor filters 410, 412, ..., 414. Is saved. Next, the average filter response value 430 is obtained and becomes a representative value for the corresponding point of the network structure.

5 shows an example of training a model image for feature point matching using a network structure.

In FIG. 5, 500 represents a normal (average) sized face image, and 502 represents a face image reduced to 1/2 size. That is, we trained by applying a rectangular mesh structure to two resolutions. The network structure used in the training process is applied to the input image in the same shape to correspond to various sizes. As illustrated in FIG. 5, when training to have two resolutions, if the average size of a face image used in training is M, a face within an M / 2 to 2 * M size range may be detected.

6 (a) to 6 (c) show matching examples for face detection.

6 (a) and 6 (b) show how faces of various sizes included in the input image match the model image database. 6A illustrates an example of face detection when the face 600 in the input image has an average size. If the face is the average size in the input image 600, the input image 602 that matches the model image 500 of the average size and is reduced to 1/2 as shown in FIG. Match 502. 6 (b) is a matching example for face detection when the face 610 in the input image has a large size (2 * average). If the face is large in the input image 610, the input image reduced by 1/2 is matched with the normal size model image 500.

6 (c) shows how the size of the network structure can be changed and applied to the actual input image. The black dots are the points from which the filter response values are calculated in the input image, and the white dots represent rectangular network structures of various sizes. As shown in FIG. 6C, the size of the network structure may be changed to match the model image database without reducing or increasing the size of the input image.

According to the above-described process, the results of the experimental images acquired in the real office environment (complex lighting and background) are as follows. The conventional first method, Bunch Graph Matching, as a comparative example, had a face detection rate of 76%, and the second method, a lattice matching method with an independent model database, had a face detection rate of 91%. On the other hand, the present invention showed the best performance with a face detection rate of 96%.

As a result of the experiment, the present invention has a smaller size of the model image database to be maintained compared to the batch graph matching method, reduces the amount of computation required for matching, and realizes more reliable face detection than the lattice matching method using a single filter mean value. The method of the present invention, which attempts to match using a network structure, overcomes the problem of misdetection that occurs when one or two feature points are incorrectly found in a bunch graph matching method that independently extracts several points and constructs a graph. Compared to other methods of applying the resolution, the same resolution range can be searched with a relatively small amount of computation.

As described above, in the face detection method and apparatus according to the present invention, first, the amount of calculation required during matching is significantly smaller than that of the bunch graph matching method or the individual matching method, and the maximum (face The computation amount by the number of feature groups in the number / number of image groups).

Second, compared with the method of finding the feature points individually through matching using a network-type search area with positional constraints among the feature points, it improves the reliability and average response value of each group for the grouped model face images rather than a simple average value. By using it, face detection performance is improved.

Third, since we construct a model database with only mean values for groups, we need to maintain a feature model database that is significantly smaller than the database for bunch graph matching or individual matching.

Fourth, the model database is trained by applying network structures of various sizes and applied to matching to flexibly respond to face size changes.

Claims (10)

In the face detection method for finding a human face in the input image, (a) generating a model image database by grouping facial images to be referred to for finding a face by feature; And and (b) detecting a face matched with the model image database from an input image by using a network-type search region corresponding to a face size to be searched for. According to claim 1, wherein the step (a), (a1) grouping facial images by features; (a2) determining a search area having a mesh shape corresponding to a face size to be searched for; (a3) obtaining filter response values for the face images in each group for each feature point in the search region, and obtaining the average filter response value as a representative value for the feature point; And and (a4) constructing a model image database with average filter response values of the grouped face images for the search region. The method of claim 2, wherein step (a) comprises: (a5) training the model image database by changing the face image at various resolutions so that an average filter response value of the face image corresponds to various face sizes. According to claim 1, wherein step (b), (b1) obtaining filter response values over the entire input image; (b2) identifying a network type search area used to obtain average filter response values of the face images grouped in step (a); (b3) comparing the filter response value of each point in the search region with the average filter response value of the corresponding feature points in the grouped face images included in the model image database while scanning the search region over the entire input image. step; And and (b4) detecting a face matching the model image database from the input image as a result of the comparison. The method of claim 4, wherein step (b3), And changing the size of the input image and / or the size of a face image in the model image database. The method of claim 4, wherein the mesh-shaped search area, Face detection method characterized in that it is adjustable to a variable size. In the face detection device for finding a face of a person in the input image, A face image reference unit for generating a model image database by grouping face images to be referred to for finding a face by feature; And And a face detector which detects a face matching the model image database from the input image by using a mesh-shaped search region corresponding to the face size to be searched for. The apparatus of claim 7, wherein the face image reference unit comprises: A grouping unit which groups and outputs face images according to features; An image filter unit for obtaining a filter response value of the face images in each group for each feature point in the search area using the search area, and outputting the average filter response value as a representative value for the feature point; And And a model image database storing a face image representing each group and average filter response values obtained by the image filter unit. The method of claim 8, wherein the face detection unit, An image filter obtaining a filter response value over the entire input image; While scanning the search region over the entire input image, the face candidate region in the input image is compared by comparing the filter response value of each point in the search region with the average filter response value of the corresponding feature point in the face image of the model image database. A first search unit for detecting; And And a second search unit for detecting a feature point representing a face in the face candidate region. The method of claim 7, wherein the search area in the form of a mesh, Face detection apparatus, characterized in that adjustable to a variable size.