KR20060020998A - Method for detecting a face and apparatus thereof - Google Patents

Abstract

The present invention relates to a method for extracting an optimal face area from a digital image such as a moving picture or a still image or an image including a human face, and in particular, an image obtained from various image acquisition tools such as a mobile phone, a digital camera, and a web camera. A method for speeding up / optimizing facial region extraction is provided.

 The present invention relates to a method for generating feature information of each divided region by dividing an input image into several small regions, and extracting a facial region using the learned results by learning the generated feature values.

; The present invention provides a preprocessing process for an input image / image, a feature value generation process for image extraction, a process of classifying pre-learned results into a face and a non-face region, a process of extracting a face, and a process of learning a face. Include.

Facial area, extraction method, AdaBoosting learning method

Description

{Method for detecting a face and apparatus}             

1 is a block diagram showing the overall configuration of the present invention.

2 is an overall flowchart of a method for face extraction.

3 is a detailed flowchart of an AdaBoosting learning method.

4 is a view of extracting a face according to the present invention.

In general, face extraction is a necessary process in face recognition process, and the feature value is generated through filtering process of face feature using normalized extracted face area and used for face recognition process. In face extraction, a method of finding a face region from face color information is used as a typical method. However, the extraction of face position by color information is highly dependent on the image acquisition conditions such as the camera or the light that acquired the image. In particular, the face or skin color value changes considerably according to the illumination, and when the image acquisition condition is not known, it is difficult to determine the range of the skin color value to determine only the face color area. In addition, the process of determining only the face region for the similar skin color extracted in a wide range including the background region has a lot of difficulties.

The present invention was devised to solve the above problems, and an object thereof is to provide a method for quickly and accurately extracting a face even under various environmental conditions when an image is acquired.

In order to achieve the above object, the face extraction method according to the present invention comprises extracting a representative value by summing each pixel value of a predetermined sub region (SubWindow) of an input image and using this value to find a contour component of the image. Feature value extracting means for extracting feature values using a filter; Learning means for learning a face using the extracted feature values; It includes a classification means for distinguishing the face and the non-face area by using the learned results.

Hereinafter, a basic concept applied to the present invention will be described.

In the present invention, unlike extracting a face through input of a limited condition such as a color image, a black and white image, and a complex background image, the region is divided into a predetermined subregion after splitting the input image with an image of an unrestricted condition. All pixel values included are summed and extracted as a representative value of the area. This process requires less computation and removes local noise than conventional pixel processing. In order to extract the feature values of the face area, the difference between the representative values is calculated in the vertical direction of the neighboring area and the diagonal direction. The feature value is used to train the face region and the non-face region through the learning means. In order to learn, we need to construct an image set with a face and an image set without a face. The image size of the data to be learned is adjusted to the same size through a process of normalization. In this case, the face image set is configured to extract a face even under various shooting environment conditions such as a face face, a rotated face, a face with different lighting conditions, and a face covering a part of the face with the back of the hand.

The learning method was AdaBoosting method. Since this method shows stronger learning performance on noise-free data, it is trained after the noise reduction filter processing on the image to be learned. The performance of the AdaBoosting method is determined by three factors:

First, when facial feature values are calculated in all sub-areas and used in stages, performance and execution time are determined by which feature value is used in the first Weak Learner. Therefore, in the present invention, using the fact that the presence of the eye in the face region has a large difference in brightness from the surrounding face color region, the feature value around the eye is learned.

Second, performance depends on how to set the threshold value going to the above step in the structure processed by cascade. Through experiments, the point where the sum of the error rate that finds the non-face area as the face and the error rate that finds the face as the non-face area is minimized.

Third, the performance depends on how many steps the successive steps are made. This also experimented to determine the point of minimum error.

By combining the three factors as above, the face extraction process is performed using the pre-learned results. Since the image size of the input image is different in the face extraction process, in the process of dividing into subregions, the first basic subregion size starts with 24 pixels horizontally and 24 pixels vertically, and then the representative value is extracted by increasing or decreasing the subregion size by a constant ratio. do. In this way, we learn to converge until the error is minimized.

3 illustrates a process of separating a face and a non-face using a learning process and a learned result. The face determined in the final step is determined to be a face using a classifier composed of steps.

The present invention detects a face stably and quickly even under various image acquisition conditions, and thus shows the same performance as a general camera even in a simple device or an environment in which image quality is poor, such as a camera attached to a PC-based web camera or a mobile phone.

The system and method according to the present invention provide an effective and optimized method in areas where facial area extraction is required, such as a computer system that understands human behavior and responds accordingly, and a surveillance system that controls access to a computer or building.

Claims (9)

The process of preprocessing for brightness correction and noise reduction of the original image, the process of dividing the preprocessed original image into a certain region, and extracting feature information of each segmented region of the segmented region. And detecting a face region as a result of the learning. (a) Raw image preprocessing (b) Raw image constant region segmentation process (c) Process of Extracting Feature Values Using Segmented Regions (d) The process of learning to use feature values (e) The process of detecting facial areas using the learned results The method of claim 2, In the process of pre-processing the image in the step (a), the histogram is adjusted to the uniform brightness region using the pixel value histogram of the image, and the method of down sampling to remove the image noise The method of claim 2, In step (b), after dividing the original image into a predetermined region, a method of obtaining a representative value of the region by adding up pixel values of the divided region The method of claim 2, In the step (c), using the representative value of the division region obtained in the step (b), the feature value is extracted by obtaining the difference between the representative value in the region located in the direction of the up, down, left, and right diagonal directions adjacent to the region to obtain the feature value. The method of claim 2, In the step (d), the method of classifying the face and the non-face area using the feature value extracted in the step (c) is performed using the AdaBoosting learning method. The method of claim 6, The images to be trained are divided into the same number of face sets and the faceless set of images, and the faced image sets are the face covered by the front face, rotated face, hands, etc. How to make up your face The method of claim 2, How to classify face areas using the results learned in step (e) The method of claim 8, Threshold is used in each stage classifier that distinguishes face and non-face using average of feature values used in each stage.

Images