KR20100030997A - Apparatus and method for extracting feature point to recognize face/object - Google Patents

Abstract

PURPOSE: An apparatus and a method for detecting feature points to recognize face/object are provided to exactly recognize the face and object although the illumination of the face or object changes and to ensure excellent recognition and processing speed. CONSTITUTION: An equalization unit(141) can obtain the image from which the illuminating effect is removed by applying pixel unit histogram equalization. A contraction unit(142) contracts the equalized image by down-sampling the image. A normalization unit(143) varies a brightness value by performing normalization to the contracted image. A vector converter(144) converts the normalized image into a vector. A controller(145) corresponds the converted vector to one point of a predetermined dimension feature space.

Description

Apparatus and method for extracting feature point to recognize face / object}

The present invention relates to a recognition apparatus and a method, and more particularly, to a feature point detection apparatus and method for face / object recognition for detecting a feature point to be able to recognize a face or an object without being affected by changes in illumination.

Face / object recognition devices are becoming more and more important in security systems and commercial applications, and their potential applications in surveillance systems and robots are growing as strategic industries in developed countries. It is an area that is highly regarded as important. However, there is a problem that the recognition performance of the face / object recognition device greatly depends on the lighting change.

Various methods have been proposed to obtain high recognition rate even under various lighting changes. The existing methods can be categorized into three categories. First, the effect of lighting is calculated in three-dimensional space using a three-dimensional model, and the second is the eye, nose, mouth, or object of the face. There are methods of extracting prominent features and removing lighting on faces or objects, and finally removing light effects using image processing techniques on secondary image images of faces or objects. The method using the 3D model is quite accurate or performance, but it is not suitable for use in the recognition system due to the large amount of computation. The method of extracting a feature of a face or an object requires less computation than a method of using a 3D model, but there is a problem in that a recognition rate is greatly reduced when a feature is incorrectly extracted or the extracted feature is out of position. Finally, the method of using 2D video images is the least computational and suitable for use in real systems, but the recognition rate is lower than that of the previous two methods.

The technical problem to be solved by the present invention is to provide a feature point detection apparatus and method for face / object recognition that can obtain a high recognition performance by solving the problem of lowering the face / object recognition rate caused by various lighting changes.

In order to solve the technical problem to be solved by the present invention, a feature point detection apparatus for face / object recognition removes an illumination effect from a received image and detects the feature point by vector converting the image from which the illumination effect is removed. It is preferable to include.

The digital signal processing means may include: a smoothing unit configured to obtain an image from which an illumination effect is removed by applying histogram smoothing on a pixel basis from a received image; A reduction unit for downsampling the smoothed image to reduce the size; A normalizer for changing a brightness value by performing normalization on the reduced image; A vector converter converting the normalized image into a vector; And a controller that corresponds the converted vector to one point of a predetermined dimension feature space.

In the present invention, the smoothing unit generates a histogram for a window having a predetermined size defined in an arbitrary pixel in a received image, obtains a cumulative probability distribution function from the histogram, and uses the cumulative probability distribution function of the pixel. After defining a transform function that creates a new brightness value, we can apply it to all pixels.

In the present invention, the smoothing unit maintains a histogram of one pixel adjacent to the arbitrary pixel, generates a new histogram by accessing only pixels in a window area different from the window, and generates a cumulative probability distribution function from the histogram. After obtaining a transformation function for generating a new brightness value of the pixel by using the cumulative probability distribution function, it may be applied to all pixels.

In the present invention, the normalization unit obtains an average and a variance of the brightness values of the pixels of the reduced image, changes the brightness value so that the average is 0 and the variance is 1, and then generates an m × n matrix. .

In the present invention, the vector conversion unit converts the m × n matrices into a vector having m × n elements, and the control unit may correspond to the image as a point of the m × n dimensional feature space.

A feature point detection method for face / object recognition for solving the technical problem to be achieved by the present invention comprises the steps of: (a) removing the lighting effect from the received image; And (b) vector converting the image from which the lighting effect is removed to detect a feature point.

In the present invention, it is possible to obtain an image from which the lighting effect is removed by applying histogram smoothing on a pixel basis from the image received in step (a).

In the present invention, the step (a) comprises the steps of: generating a histogram for a window of a predetermined size defined in an arbitrary pixel in the received image; Obtaining a cumulative probability distribution function from the histogram; Defining a transform function for generating a new brightness value of the pixel using the cumulative probability distribution function; And applying the process to all pixels.

In the present invention, the step (a) comprises: generating a new histogram by accessing only pixels in a window area different from the window while maintaining a histogram of one pixel adjacent to the arbitrary pixel; Obtaining a cumulative probability distribution function from the histogram; Defining a transform function for generating a new brightness value of the pixel using the cumulative probability distribution function; And applying the process to all pixels.

In the present invention, the step (b) comprises the steps of (b-1) downsampling the image from which the lighting effect is removed; (b-2) changing the brightness value by performing normalization on the reduced image; And (b-3) converting the normalized image into a vector and corresponding to a point of a predetermined dimensional feature space.

In the present invention, in the step (b-2), after calculating the average and variance of the brightness values of the pixels of the reduced image, the brightness value is changed so that the average is 0 and the variance is 1, and then the m × n matrix is obtained. Can be generated.

In the present invention, step (b-3) may convert the m × n matrices into a vector having m × n elements and map the image to a point in an m × n-dimensional feature space.

According to the present invention, the face and the object can be accurately recognized even if the illumination of the face or the object is changed, and the recognition performance and processing speed are much higher than other methods. In particular, it can be easily applied to other systems such as a robotic system or a surveillance system because of its excellent recognition performance even in the case of severe lighting changes.

1 is a block diagram showing the configuration of a feature point detection apparatus for face / object recognition according to the present invention, the image input means 110, the storage means 120, the recognition means 130 and the digital signal processing means 140 It includes.

The image input unit 110 receives an image for detecting a feature point. For example, the image received by the image input unit 110 is illustrated in FIG. 2A.

The storage means 120 stores the image received by the image input means 110, and stores the feature point detection value including various data generated by the digital signal processing means 140.

The recognition unit 130 determines whose face is the feature point for the face and which object is the feature point for the object based on the feature point detected by the digital signal processing unit 140.

The digital signal processing unit 140 removes the lighting effect from the image received through the image input unit 110 and detects the feature point by vector converting the image from which the lighting effect has been removed.

To this end, the digital signal processing unit 140 includes a smoothing unit 141, a reduction unit 142, a normalization unit 143, a vector converter 144, and a controller 145.

The smoothing unit 141 generates an image from which the lighting effect is removed by applying pixel-wise histogram equalization from the received image under the control of the controller 145.

Five steps of smoothing the pixel histogram of the smoothing unit 141 are performed. A histogram is generated for a defined w × w window, a cumulative distribution function is obtained from the histogram, and a cumulative probability distribution function is obtained. The conversion function is defined, the new brightness value is determined using the conversion function, and the process is applied to all pixels of the image.

As shown in FIG. 3, a w × w window having an arbitrary size is defined for one pixel P (x, y) in an image, and the cumulative probability distribution function C (i) as shown in Equation 1 is defined as pixels belonging to the window. Obtain

In Equation 1, p (j) means a brightness value of each pixel belonging to the window. The cumulative probability distribution function C (i) is used to define a transform function i _he as shown in Equation (2).

In Equation 2, max and min mean maximum and minimum brightness values of each pixel. Calculate the new brightness value i _he (x, y) of one pixel P (x, y) using this conversion function i _he , and repeat this process for every pixel in the image. Can be obtained. FIG. 2B illustrates an example of an image generated by the pixel histogram smoothing process of the smoothing unit 141 with respect to the input image of FIG. 2A, that is, an image in which the lighting effect is removed.

When the size of the window is w × w, to perform pixel-by-pixel histogram smoothing for each pixel, we need to access w × w pixels. However, when performing the pixels, histogram smoothing for every pixel in the image can be accessed as a single w ² pixels perform histogram equalization only access single 2w.

As shown in FIG. 4, if the pixel histogram smoothing is performed for one pixel P (x, y) in the image, the histogram of the window centered on the pixel P (x, y) is known. When performing pixel-by-pixel histogram smoothing on pixels P (x, y) and neighboring pixels P '(x + 1, y), we know that we know the histogram of the window centered on pixels P (x, y). By subtracting the information of pixels corresponding to R _O of the window centered on the pixel P (x, y) and adding the information of the pixels corresponding to R _n , the window centered on the pixel P '(x + 1, y) A histogram of can be obtained. If the histogram of the neighboring pixels is still maintained when performing the pixel histogram smoothing, it is possible to perform the pixel histogram smoothing at a high speed because a new histogram can be obtained by only approaching ² w pixels instead of w ² . .

Under the control of the controller 145, the reduction unit 142 performs a histogram smoothing on a pixel basis to downsample and reduce the image from which the lighting effect is removed.

Using the original M × N image as it is is too large and requires a lot of memory and computation. Therefore, the dimension of the data should be reduced, and the reduction unit 142 reduces the size of the image to a desired size of m × n (M >> m, N >> n) by using a down sampling method that is frequently used in image processing. . Other methods, such as principal component analysis (PCA) and linear discriminant analysis (LDA), can be used to reduce image size while minimizing data distortion. Principal component analysis or linear discriminant analysis requires a large amount of memory and calculation amount, so in the present invention, the downsampling method is used to reduce the image simply and quickly. FIG. 2C illustrates an example in which a histogram smoothing by a pixel is performed to reduce an image from which an illumination effect is removed.

The normalizer 143 changes the brightness value by performing normalization on the reduced image under the control of the controller 145.

The normalization unit 143 obtains the mean and the variance of the brightness values for each pixel of the reduced m × n size image, and then changes the brightness values of all the points such that the average is 0 and the variance is 1. do. An m × n matrix may be obtained through the image normalization process of the normalization unit 143, and a matrix generated through the image normalization process is illustrated in FIG. 2D.

The vector converter 144 converts the normalized image into a vector under the control of the controller 145, and converts the m × n matrix into a vector having m × n elements. 2E shows an example of converting an m × n matrix into a vector having m × n elements.

The controller 145 corresponds a vector to a point of m × n dimension, an example of which is shown in FIG. 2F. The controller 145 maps the M × N size 2D image to a point of the m × n dimension feature space, wherein the images of the same person acquired under different lighting conditions are located at the same point in the feature space, and Images of different people are located at different points in the feature space. This eliminates all changes due to lighting effects and places them in the feature space according to differences in inherent shapes of faces or objects, making it possible to create a recognition device that is robust to changes in lighting. The transformed m × n dimensional vectors can be easily classified into individual people or objects using various existing classifiers. In FIG. 5, when the image of the face is converted, the image image of the face obtained under different lighting conditions for the same person corresponds to the same position in the multi-dimensional space, and the image image of the face obtained under the same lighting condition for different characters is different in the multi-dimensional space. The embodiment illustrates that different people can be distinguished from each other by corresponding to the location.

Next, a feature point detection method for face / object recognition according to the present invention will be described with reference to FIGS. 6 and 7. The feature point detection method for face / object recognition according to the present invention may be performed inside the feature point detection device for face / object recognition as shown in FIG. 1. With the help of the components it can be performed inside the tracking digital signal processing means 140.

FIG. 6 is a flowchart illustrating an operation of a method for detecting a feature point for face / object recognition according to the present invention. When the digital signal processor 140 receives an image (step 610), the pixel effect histogram smoothing is applied to the image. Acquire the removed image (step 620).

FIG. 7 illustrates a detailed method according to the first embodiment of the pixel-to-pixel histogram smoothing process applied by the digital signal processor 140.

First, the digital signal processor 140 defines an arbitrary w × w in an image and generates a histogram from the defined w × w window (step 621). Since histogram generation has many known technologies, description thereof will be omitted.

When the histogram is generated, the digital signal processor 140 obtains a cumulative probability distribution function from the histogram (step 622). As shown in FIG. 3, the digital signal processor 140 defines a w × w window having an arbitrary size for one pixel P (x, y) in an image, and the pixels belonging to the window are represented by Equation 1 above. Obtain the cumulative probability distribution function C (i). In Equation 1, p (j) means a brightness value of each pixel belonging to the window.

When the cumulative probability distribution function is obtained, the digital signal processor 140 defines a transform function using the cumulative probability distribution function (step 623). The digital signal processor 140 defines a transform function i _he as shown in Equation 2 using the cumulative probability distribution function C (i). In Equation 2, max and min mean maximum and minimum brightness values of each pixel.

When the transform function is defined, the digital signal processor 140 determines the new brightness value i _he (x, y) of the arbitrary pixel P (x, y) using the transform function (step 624), and processes the image. When applied to all pixels of the control panel (step 625), an image from which the lighting effect is removed may be obtained.

When the size of the window is w × w, to perform pixel-by-pixel histogram smoothing for each pixel, we need to access w × w pixels. However, when performing the pixels, histogram smoothing for every pixel in the image can be accessed as a single w ² pixels perform histogram equalization only access single 2w. As shown in FIG. 4, if the pixel histogram smoothing is performed for one pixel P (x, y) in the image, the histogram of the window centered on the pixel P (x, y) is known. When performing pixel-by-pixel histogram smoothing on pixels P (x, y) and neighboring pixels P '(x + 1, y), we know that we know the histogram of the window centered on pixels P (x, y). By subtracting the information of pixels corresponding to R _O of the window centered on the pixel P (x, y) and adding the information of the pixels corresponding to R _n , the window centered on the pixel P '(x + 1, y) A histogram of can be obtained. If the histogram of the neighboring pixels is still maintained when performing the pixel histogram smoothing, the pixel histogram smoothing can be performed at high speed because a new histogram can be obtained by only approaching ² w pixels instead of w ² .

6, when the image in which the lighting effect is removed is obtained, the digital signal processor 140 reduces the size of the image (step 630).

Using the original M × N image as it is is too large and requires a lot of memory and computation. Therefore, the dimension of the data needs to be reduced. The digital signal processor 140 reduces the size of the image to a desired size of m × n (M >> m, N >> n) by using a down sampling method that is frequently used in image processing. do.

When the image is reduced, the digital signal processor 140 normalizes the reduced image and converts the image into a matrix (step 640).

The digital signal processor 140 obtains the mean and the variance of the brightness values for each pixel of the reduced size of m × n, and then adjusts the brightness values of all points such that the average is 0 and the variance is 1. Change it. An m × n matrix may be obtained through the image normalization process of the digital signal processor 140.

When the matrix is obtained, the digital signal processor 140 converts the matrix of size m × n into a vector having m × n elements (step 650).

When the vector conversion is completed, the digital signal processor 140 associates the vector with a point of a predetermined (m × n) dimensional feature space (step 660).

The two-dimensional image of the M × N size is mapped to a point in the m × n-dimensional feature space, and the images of the same person acquired under different lighting conditions are located at the same point in the feature space. Position it at a different point in the feature space. This eliminates all changes due to lighting effects and places them in the feature space according to differences in inherent shapes of faces or objects, making it possible to create a recognition device that is robust to changes in lighting. The transformed m × n dimensional vectors can be easily classified into individual people or objects using various existing classifiers.

1 is a block diagram showing the configuration of a feature point detection device for face / object recognition according to the present invention.

FIG. 2 is a diagram illustrating an operation process of a feature point detection block for face / object recognition in the apparatus of FIG. 1.

FIG. 3 is a diagram for describing a first embodiment of pixel-to-pixel histogram smoothing in the apparatus of FIG. 1.

FIG. 4 is a diagram for describing a second exemplary embodiment of pixel- histogram smoothing in the apparatus of FIG. 1.

FIG. 5 is a diagram illustrating an example of an image obtained in different lighting conditions for the same person and an image acquired under the same lighting condition for another person in a device in FIG. 1.

6 is a flowchart illustrating an operation of a method for detecting a feature point for face / object recognition according to the present invention.

FIG. 7 is a flowchart illustrating a method of obtaining an image in which lighting effects are removed by applying a histogram smoothing unit by pixel in FIG. 6.

Claims (13)

And digital signal processing means for removing the lighting effect from the received image and vector transforming the image from which the lighting effect has been removed to detect the characteristic point. The method of claim 1, wherein the digital signal processing means A smoothing unit configured to obtain an image from which the lighting effect is removed by applying a histogram smoothing pixel by pixel from the received image; A reduction unit for downsampling the smoothed image to reduce the size; A normalizer for changing a brightness value by performing normalization on the reduced image; A vector converter converting the normalized image into a vector; And And a controller for mapping the transformed vector to one point of a predetermined dimensional feature space. The method of claim 2, wherein the smoothing unit A transform function that generates a histogram for a window of a predetermined size defined in a given pixel in a received image, obtains a cumulative probability distribution function from the histogram, and generates a new brightness value of the pixel using the cumulative probability distribution function. The apparatus for detecting a feature point for face / object recognition according to claim 1, wherein the image is defined and then applied to all pixels. The method of claim 3, wherein the smoothing unit While maintaining a histogram of one pixel neighboring the arbitrary pixel, a new histogram is generated by accessing only pixels in a window area different from the window, a cumulative probability distribution function is obtained from the histogram, and the cumulative probability distribution function is calculated. And a conversion function for generating a new brightness value of the pixel by using the same, and applying the same to all pixels. The method of claim 3, wherein the normalization unit After obtaining the average and the variance of the brightness value of each pixel of the reduced image, the brightness value is changed so that the average is 0, the variance is 1, and then generates an m × n matrix for face / object recognition Feature point detection device. The method of claim 5, The vector converter Converting the m × n matrices into a vector having m × n elements, The control unit And the image is mapped to a point in an m × n dimensional feature space. (a) removing the lighting effect from the received image; And (b) detecting a feature point by vector transforming the image from which the lighting effect has been removed. 8. The method of claim 7, wherein in step (a) A method for detecting a feature point for face / object recognition, comprising: applying a histogram smoothing on a pixel basis from a received image to obtain an image from which an illumination effect is removed. The method of claim 8, wherein step (a) Generating a histogram for a window of a predetermined size defined for any pixel in the received image; Obtaining a cumulative probability distribution function from the histogram; Defining a transform function for generating a new brightness value of the pixel using the cumulative probability distribution function; And And applying the process to all the pixels. The method of claim 9, wherein step (a) Generating a new histogram by accessing only pixels in a window area different from the window while maintaining a histogram of one pixel neighboring the arbitrary pixel; Obtaining a cumulative probability distribution function from the histogram; Defining a transform function for generating a new brightness value of the pixel using the cumulative probability distribution function; And And applying the process to all the pixels. The method of claim 7, wherein step (b) (b-1) downsampling and contracting the image from which the illumination effect is removed; (b-2) changing the brightness value by performing normalization on the reduced image; And (b-3) converting the normalized image into a vector and corresponding to a point in a predetermined dimensional feature space. The method of claim 11, wherein step (b-2) After obtaining the average and the variance of the brightness values of the pixels of the reduced image, changing the brightness value so that the average is 0 and the variance is 1, and then generating an m × n matrix for face / object recognition. Feature point detection method. The method of claim 12, wherein step (b-3) And converting the m × n matrices into a vector having m × n elements and correlating the image to a point in an m × n dimensional feature space.

Images