KR101143555B1 - Face Detection System using Skin Color Filtering and Morphology Process and Method Therefor - Google Patents

Abstract

According to the present invention, by performing face color filtering and morphology calculation on an image in a face detection preprocessing step for selecting a candidate face region in a color image, a calculation consumption and background for detecting a face in an unnecessary region are determined. Face false detection can be reduced. In addition, by rotating the input image in a clockwise or counterclockwise direction and performing face detection on each of the rotated images, the inclined face region existing in the image can be detected.
According to an aspect of the present invention, there is provided a face detection method comprising: (a) a preprocessing step of correcting distortion of an input image by size and illumination; (b) if the color model is not YCbCr, converting to a YCbCr color model; (c) performing face color filtering using color difference information of Cb and Cr; (d) a morphology calculation step of removing noise of an image on which color filtering is performed; (e) selecting a candidate face image in a non-black region from the result of step (d); And (f) detecting a front face from the candidate face image by using a face candidate search window.

Description

Face Detection System using Skin Color Filtering and Morphology Process and Method Therefor}

The present invention relates to a face detection system and method, and more particularly, to a face detection system and method using face color filtering and morphology calculation.

In the field of image processing, face detection is defined as finding the position and size of a face existing in image data. Face detection technology is closely related to face recognition technology and deeply related to various applications. However, since the face of a person present in the image data appears in various forms due to factors such as background, posture, facial expression, and lighting change, it is difficult to accurately detect only the face of the person from the image.

The face detection methods that have been studied so far can be classified into the following forms.

First is a knowledge-based method. This is a method of detecting a face using the assumption that each component (eye, nose, mouth, etc.) exists on a face of a person, and each element has a certain position and distance relationship. Such a method is relatively easy to find a face by using the components of the face in the normalized image composed of the front, but face detection is difficult for various changes, such as the tilt, angle, facial expression of the face.

The second method is to use information of facial feature components. This is a method of detecting a face by using mixed information such as size, shape, color, and correlation of each component of facial feature components. This may have the advantage of easy face detection and fast processing time by utilizing a variety of information, but there is a problem that is sensitive to noise due to changes in illumination, degree of tilt of the face, posture, and complex background.

Third, the method uses template matching. Template matching is a method of detecting a face from an image by generating a standard template for all the faces to be detected and comparing it with an input image. Such a method has a more robust advantage in the background and lighting because there is no need to extract the presence of a feature component for detecting a face and its features as in the aforementioned method. However, it is sensitive to distortions such as the change in the size of the face image present in the image, the tilt and rotation of the face according to the direction of the gaze. In addition, it requires a lot of effort to create a standard template for the faces of people with various expressions, postures, and appearances.

Fourth, it is an appearance-based method. The appearance-based method is a method of detecting a face using a model trained by a set of training images. To do this, we create a classifier or classification model that can classify faces and non-faces using a training image set composed of face and non-face images, and find a face in the image. Various methods such as PCA, LDA and SVM exist, and these methods show relatively good performance because they can overcome various constraints of the other detection methods introduced above through model training. However, the training of the classifier takes a lot of time, and since the classifier is affected by the training data, the model has to be retrained when the database changes.

Most studies on face detection use only grayscale images and do not use color information. However, the color feature is one of important features in distinguishing the face from the non-face, and if the face and the non-face can be distinguished using the color information, more efficient face detection can be performed. In addition, the above-described methods have a problem in that a large amount of calculation is required to increase the detection accuracy when the size of an image is large or there are a large number of people in the image, as well as a false detection rate.

Disclosure of Invention The present invention has an object to solve the above technical problem, and is an unnecessary area by performing face color filtering and morphology calculation in a face detection preprocessing step for selecting a candidate face area in a color image. The purpose of the present invention is to reduce the computational consumption for face detection and the detection of face false detection by background.

It is also an object to detect the inclined face region existing in the image by rotating the input image clockwise and counterclockwise, and performing face detection on each rotated image.

According to an aspect of the present invention, there is provided a face detection method using face color filtering and morphology calculation, the method comprising: (a) a preprocessing step of correcting distortion of an input image by size and illumination; (b) if the color model is not YCbCr, converting to a YCbCr color model; (c) performing face color filtering using color difference information of Cb and Cr; (d) a morphology calculation step of removing noise of an image on which color filtering is performed; (e) selecting a non-black region as a candidate face image from the result of step (d); And (f) detecting a face from the candidate face image using a face candidate search window.

In addition, the step (a) preferably comprises reducing the size of the input image to the size of the set resolution, if the size of the input image is larger than the preset resolution.

Specifically, the step (a) further comprises performing at least one of gamma correction and histogram equalization to reduce distortion caused by changes in illumination of the input image.

Specifically, the morphology operation of the step (d) performs smoothing and dilation of the color-filtered image and smoothes the contours, and removes the hole generated in the color filtering process. Infill is characterized by a role.

The method of performing the contraction and expansion operation is an opening operation in which small areas belonging to the background are removed by performing the expansion operation after the shrinkage operation.

In addition, the method according to another embodiment of performing the contraction and expansion operation is characterized in that the closing operation to remove the noise isolated in the face area by performing the contraction operation after the expansion operation.

In the method (f), the face candidate search window searches for face detection by scanning a subregion having a constant size in the image with respect to a portion selected as the candidate face image through color filtering and morphology calculation. And by determining if a face exists.

In addition, according to an embodiment of the present invention, after the step (f), (g) excluding the face area detected in the step (f) from the clockwise search candidate; (h) rotating the candidate face image in a clockwise direction; (i) resuming the search for detecting the rotated face in the image to detect the face; (j) excluding a face region detected in step (i) from a counterclockwise search candidate; (k) rotating the candidate face image counterclockwise; (l) resuming the search for detecting the rotated face in the image to detect the face; And (m) detecting the final face by combining the faces detected in steps (f), (i) and (l).

According to the face detection method using the face color filtering and the morphology operation of the present invention, the face detection in the unnecessary area is performed by performing the face color filtering and the morphology operation in the face detection preprocessing step for selecting the candidate face area in the color image. It is possible to reduce the amount of computation required and false detection of the face by the background.

In addition, by rotating the input image in a clockwise or counterclockwise direction and performing face detection on each of the rotated images, the inclined face region existing in the image can be detected.

1 shows a face detection system according to a preferred embodiment of the present invention.
2 is a flowchart illustrating a face detection method according to an exemplary embodiment of the present invention.
3 illustrates an input image used for an embodiment of the present invention.
4 and 5 illustrate processing results of an input image according to execution of each step of the flowchart of FIG. 2.
6 shows the degree of conversion of the image data by the gamma value.
7 shows one embodiment of histogram equalization.
8 illustrates one embodiment of erosion and expansion operations.

Hereinafter, a face detection system and method using face color filtering and morphology calculation according to an embodiment of the present invention will be described in detail with reference to the accompanying drawings.

The following examples of the present invention are intended to embody the present invention, but not to limit or limit the scope of the present invention. It will be understood by those of ordinary skill in the art that various changes in form and details may be made therein without departing from the spirit and scope of the invention as defined by the appended claims.

1 shows a face detection system according to a preferred embodiment of the present invention.

As can be seen from FIG. 1, the face detection system 100 according to an exemplary embodiment of the present invention includes an image input unit 10, an image processing unit 20, and an image output unit 30.

The image processor 20 may include an image preprocessor 21, a color model converter 22, a filter 23, a morphology calculator 24, a candidate face image selector 25, a face detector 26, and a face. The storage unit 27 and the image rotating unit 28 is included.

An image to be used in the system is input using the image input unit 10 such as a digital camera or a camcorder, and the input image is stored in the face storage unit 27 to enable processing in the system.

That is, the image processing unit 20 detects an image and the face of the image input to the image input unit 10, and stores the image detected by the image output unit 30 to a display, a printer, and an external device. It serves to output through.

2, a face detection method according to an exemplary embodiment of the present invention using the face detection system of the present invention will be described in more detail.

First, FIG. 3 shows an input image used for an embodiment of the present invention.

4 and 5 show the processing result of the input image according to the execution of each step on the flowchart of FIG. 2.

The image preprocessor 21 performs preprocessing S5 to correct distortion of the input image by size and illumination.

The step S5 includes reducing the size of the input image to the size of the set resolution if the size of the input image is larger than the preset resolution.

The method may further include performing at least one of gamma correction and histogram equalization to reduce distortion caused by a change in illumination of the input image.

When the size of the input image is larger than the resolution of any size set by the system, the size of the input image is reduced to 1 / N, thereby reducing the amount of computation and reducing noise.

In addition, the gamma correction will be described in more detail.

, gamma) 값으로 그 특성이 결정된다. In general, there is a difference in the display for outputting the image data received by the image storage device again. This change is different from the actual image because the brightness of the image is changed to bright and dark. Therefore, the method of compensating for the difference in brightness is gamma correction, and the gamma correction nonlinear transfer function is generally gamma (1)

, the characteristic is determined by the value gamma).

here Ci means color pixels in the input image, and C means color pixels in the gamma corrected image. The conversion degree of data by the gamma value is shown in FIG. 6. That is, when the gamma value is smaller than 1 from FIG. 6, the color of the image becomes clearer by the above equation, and when the gamma value is larger than 1, the image becomes brighter.

The histogram equalization is described in more detail below.

The human eye is more sensitive to relative brightness than absolute brightness. Therefore, even if the same image is wider than the narrowly gathered image on the histogram, the image is much clearer and more visible to the human eye. The histogram equalization algorithm using this characteristic of the human eye is an algorithm that produces a clearer image by increasing the brightness range of an image having a narrow brightness range to the maximum. A histogram of an image may be referred to as a discrete function having a range [0, L-1] and may be represented by Equation 2. Where L is the value of 2 ^B in each color channel, and B is the number of bits per pixel.

Here, r _k denotes the kth gray level in each color channel, and n _k denotes the number of pixels having the gray level r _k in the image. If the probability of occurrence of the gray level r _k is approximately represented by Equation 3, the Discrete version cdf is expressed by Equation 4.

Histogram equalization is performed through a function that maps pixels with r _k levels in the original image to the values of s _k level × (L-1) according to Equation 4 and consequently after histogram equalization. The histogram cdf is straight. The histogram equalization changes the dark, bright, and gray-colored blurry images we saw earlier into sharper contrasts, resulting in better viewing of the human eye, which increases the range of finer contrast. You can get an image.

, gamma) 값을 1로 선택하거나, 히스토그램 균등화 과정에서 수학식 3 대신

로 선택하고, 그에 맞는 상기의 ((s_k 레벨×L)-1)의 값으로 맵핑함으로써, 감마 보정 및 히스토그램 균등화 과정의 수행 여부를 결정할 수 있다. In this system, gamma correction is performed during gamma correction and histogram equalization.

, gamma) as 1, or instead of Equation 3 during histogram equalization

It is possible to determine whether to perform a gamma correction and histogram equalization process by selecting and by mapping to the value of ((s _k level × L) −1).

In other words, in the present system, it is possible to selectively perform gamma correction and histogram equalization, respectively.

7 shows an example of the original image and histogram equalization.

After the preprocessing process of step S5, the input image of the RGB color model is converted by the color model converter 22 into an image having a YCbCr color model (S10).

The general RGB color model used in the image processing field is sensitive to light changes. Therefore, the distribution of color is changed due to the influence of external lighting change. Accordingly, in the present invention, face color filtering is performed using a YCbCr model having a relatively strong characteristic against various lighting changes. Where Y represents a luminance component and Cb and Cr represent a color difference component. The advantage of using the YCbCr color model is that it is possible to use color feature information from which the influence of the luminance value is removed from the color information, thereby being able to extract a feature that is relatively robust to lighting changes.

Equation 5 below is used to convert the RGB color of the input image into a model of the color space of YCbCr.

로 설정하였고, 이때

는 각각 임의의 상수 값으로 학습 데이터로부터 사용자가 실험 환경에 의하여 임의로 설정할 수 있다.Face color filtering is performed by the filtering unit 23 using Equation 6 using color difference information of Cb and Cr of I (x, y) in the space converted to YCbCr (S15). In other words, the area that falls within the range of the face color is mapped to '255', and the area that is not within the range of the color is mapped to '0', thereby removing the non-face area that does not include the face color area from the input image. In this case, T corresponds to the range of face colors for Cb and Cr, and T is used to use the face candidate boundary color that is robust to face colors of various races.

, Where

Each can be arbitrarily set by the user in the experiment environment from the learning data as arbitrary constant values.

As a next step, a binary image subjected to face color filtering includes a lot of noises, and in order to remove such noises, a morphology calculation by the morphology operator 24 is required (S20). It performs the role of smoothing the contour and filling the hole generated in the color filtering process by performing the shrink and dilation operation on the filtered candidate face region through color filtering. In other words, the loss information generated in the pixels included in the face color region is filled with respect to the region filtered by the face candidate image through the morphology operation, and noise components generated in the background region may be removed.

8 illustrates an embodiment of a contraction operation and an expansion operation.

The opening operation, which is generally used in the morphology operation, is performed by the expansion operation after the shrink operation, so that small areas belonging to the background are removed. Closing operations, on the other hand, have the effect of eliminating noise that is isolated in the object by performing shrink operations after expansion. In the present invention, the face area corresponds to the object.

The process of noise cancellation will be described in more detail. When color filtering is performed, black areas such as eyes, nose and mouth disappear in the face area, which are filled by morphology calculation. That is, morphology calculation is used to compensate for the information lost through color filtering.

In order to perform the morphology operation as described above, an image file for performing color filtering is generated separately, and the pixel value of the corresponding position to be filled after the morphology operation is brought back from the original image of the face storage unit 27. do.

By adding the above morphology calculation to the present invention, it is possible to prevent the areas such as the eyes, the nose, the mouth, etc. from being excluded from the face area, thereby making it possible to detect the face area more accurately.

The candidate face image selecting unit 25 selects a non-black region as a candidate face image from the result of the morphology calculation (S25).

Next, the face detection unit 26 determines whether a face exists in the subregion while scanning a subregion having a predetermined size in the image using the face candidate search window. The size of the subregion may be gradually set to a large size at a predetermined interval from a predetermined small size.

In this case, an area searched by the face candidate search window for face detection is limited to the candidate face image selected by the candidate face image selecting unit 25 through color filtering and morphology calculation. By limiting the search area, the computational complexity of the scanning process of the entire area of the input image can be reduced, and the frequency of false detection in which a face is incorrectly detected in the background area can be reduced. In this manner, the face area is detected by the face detector 26 (S30). The detected image is stored in the face storage unit 27.

The face area detected in step S30 is excluded from the clockwise face search candidate area by the face detector 26 (S35). By doing so, it is possible to prevent duplicate detection of the face candidate region.

Next, after the image rotating unit 28 rotates the face candidate image clockwise, the face detecting unit 26 restarts the search to detect a face in the image (S40 and S45). The face storage unit 27 also stores the detected face image.

The face detector 26 excludes the face region detected by the clockwise rotation region (S50). Next, after the face rotating unit 28 rotates the face candidate image counterclockwise, the face detecting unit 26 starts searching again to detect a face in the image (S55 and S60). The face storage unit 27 also stores the detected face image.

As such, the face storage unit 27 stores the face regions extracted from the face candidate images rotated in the front, clockwise, and counterclockwise directions, and the final face is combined by the face storage unit 27. It extracts to a detection area (S65).

As described above, the reason for excluding the areas previously detected in the steps S35 and S50 is to prevent the face search window from searching for the area already found. The method of excluding the detection area is by erasing the corresponding part of the image in black.

In addition, by detecting the face by the clockwise and counterclockwise rotation, the detection rate of the face is increased not only for the existing front face detection but also for the face images of various angles that may occur in tilted or rotated daily life. It is possible. In addition, the clock angle and the counterclockwise rotation angle and the number of rotations can be set arbitrarily or automatically. By setting the rotation angle and the number of rotations described above, a face can be rotated clockwise or counterclockwise at regular angular intervals, and detection can be performed even for a face rotated in all 360 degrees.

10: image input unit 20: image processing unit
30: video output unit
21: image preprocessor 22: color model conversion unit
23: filtering unit 24: morphology calculator
25: candidate face image selection unit 26: face detection unit
27: face storage unit 28: the image rotating unit

Claims (12)

In the face detection method using a face detection system,
(a) a preprocessing step for correcting distortion caused by the size and illumination of the input image;
(b) if the color model is not YCbCr, converting to a YCbCr color model;
(c) performing face color filtering using color difference information of Cb and Cr;
(d) a morphology calculation step of removing noise of an image on which color filtering is performed;
(e) selecting a non-black region as a candidate face image from the result of step (d); And
(f) detecting a face from the candidate face image using a face candidate search window; Including,
When the color model of the input image is RGB in the step (b), converting to the YCbCr color model is a face detection method using the face color filtering and morphology calculation, characterized in that the following equation.

According to claim 1, wherein the step (a),
If the size of the input image is larger than the preset resolution, reducing the size of the input image to the size of the set resolution; Face detection using the face color filtering and morphology calculations comprising a.
The method of claim 2, wherein step (a) comprises:
And performing at least one of gamma correction and histogram equalization to reduce distortion caused by a change in the illumination of the input image.
delete The face detection method according to claim 1, wherein the color filtering method of step (c) is based on the following equation.

(I '(x, y) is a pixel value of an image after color filtering, I (x, y) is a pixel value of an image before color filtering, and T is a face candidate boundary color that is robust to face colors of various races. In order to

, Where

Can be set arbitrarily.)
The method of claim 1, wherein the morphology operation of the step (d),
Face color filtering and morphology calculation is performed by smoothing the contours by filling the color-filtered image by performing shrink and expansion operations, and filling the holes generated in the color filtering process. Face detection method used.
The method of claim 6, wherein the contracting and expanding operation is performed.
A face detection method using face color filtering and morphology calculation, which is an opening operation in which small areas belonging to a background are removed by performing an expansion operation after a contraction operation.
The method of claim 6, wherein the contracting and expanding operation is performed.
A face detection method using face color filtering and morphology calculation, which is a closing operation to remove noise isolated in a face region by performing a contraction operation after an expansion operation.
The method of claim 1,
In the method (f), the face candidate search window searches for face detection by scanning a subregion having a constant size in the image with respect to a portion selected as the candidate face image through color filtering and morphology calculation. A face detection method using face color filtering and morphology calculation, which is performed by determining whether a face exists.
The method of claim 1, wherein after step (f),
(g) excluding the face region detected in step (f) from the clockwise search candidate;
(h) rotating the candidate face image in a clockwise direction; And
(i) resuming the search for detecting the rotated face in the image to detect the face; the face detection method using face color filtering and morphology calculation.
The method of claim 10, wherein after step (i),
(j) excluding a face region detected in step (i) from a counterclockwise search candidate;
(k) rotating the candidate face image counterclockwise; And
(l) resuming the search for detecting the rotated face in the image to detect the face; the face detection method using face color filtering and morphology calculation, further comprising: detecting the face.
The method of claim 11, wherein after step (l),
(m) detecting the final face by combining the faces detected in steps (f), (i), and (l); and using face color filtering and morphology calculations. Face detection method.

Images