On The Facial Expression Recognition Based On The Combining Feature

Facial expression plays an important role in daily life, it is a kind of important body language that people carried on the nonverbal communication, and is the important supplement that the mankind exchanged language. The expression recognition it has been the important research direction in artificial intelligence and the field of the computer vision all the time for a long time.In recent years, with the increase of people's interest in interactive the expression recognition that becomes a research focus gradually. The expression recognition is the essential of the interactive, which is intelligence and natural. Realizing naturally harmonious Interactive, we must enable the computer to understand people's emotion and intention effectively. The expression recognition is the foundation of the emotion understanding, and it is the prerequisites and effective way of the computer understanding emotion of people. The expression recognition it is the problem that an intelligence machine comes into people's daily life and must be solved, realize the problem that the intelligence machine must face, it is people that explore intelligence and understand that the effective way.The appearance of the expression must cause the form of people's face to change. Face deformation includes the information of the expression classification. This paper denote these deforming by extracting the geometry structure feature which reflect the deforming of the face and the local statistic feature which reflect the changing of the gray level, and combine these feature to recognize the facial expression. There are four steps in this progress. Fist, preprocess the image by extracting the binary edge image. secondly, locate the eye and mouth area which have great attribute to the facial expression on the base of the BEI. Thirdly, extract the geometry structure feature and the local statistic feature by arithmetic operators and gray level co-occurrence matrix. At last, import the mix feature vector to the BP neural network, and classify the facial expression. Binary edge image extracting, feature area locating and the feature vector generating is the main point and the innovation in the paper.1) Binary edge image extractingA binary image is very important to the analyzing face image, especially the locating of the feature organ area. The experiment result show that the method of dynamic threshold based the Fisher rule function is sensitive to the sunshine. The binary image would generate the effect of gray conglutination, when the sunshine is not uniform. To avoid this effect, we use the Sobel arithmetic operators to extract the edge of the image, and generate the BEI by binary operating. There is much noise contained in the image and the image edge is not clear. To develop the BEI, we combine the binary image from the Fisher rule function and the BEI from the Sobel arithmetic operators to do the operation of logical, and obtain the clear BEI. To connect the breakpoint and fill the blank which we didn't expect, we do the operation of morphologic. The experiment result show that we can obtain the BEI quickly using the method in this paper. The BEI can conquer the effect of gray conglutination from the Fisher rule function and the noise from Sobel arithmetic operators, and can satisfy the request of the feature area locating.2) Feature area locatingThe appearance of the expression if forced by the deformation of eye and mouth these feature organ. Locating these feature area can not only conquer the"scale question"result from the large number data, but also eliminate the individual feature and reserve the useful expression information. Especially in the progress of local statistic feature extracting with the gray level co-occurrence matrix as its core algorithm, the effect of locating directly result in the feature extracting.In this paper, we use the methods of gray projection and model matching to operate the image from the JAFFE database. The experiment result show that the method of gray projection would create the question, such as:"tableland stationary point"and projection curve abnormality, because of the gray level and Hairstyle factor, result in the locating invalidation. The method of model matching is sensitive to the sunshine, and result in the vertical locating invalidation, too. So we construct a"probe"to split in the BEI using the effective data which result from the model matching and gray projection and regulate the value of the pixel. Locate the feature area from the track of probe base on the structure rule of face. The experiment result show that, using the method of"probe"to locate the feature area of the image from JAFFE database, we can arrive at an accuracy of 100% extent.3) Feature extractingFeature extracting is the most important steps which result in the classification rate of the facial expression. Nowadays method of face recognition was used in some facial expression recognition, which generates the dependence to the neutral face. To conquer this disadvantage, we propose the geometry structure feature to figure the facial expression. Using the method of Susan arithmetic operators and the edge point detection spot the feature point, after computing generate a 9-dimension feature vector to figure the deformation. To avoid the loss of the information, we generate local statistic feature using co-occurrence matrix as the supplement, more vividly describe the relationship amount the pixel and the statistic feature. The result of experiment show that the mix feature can conquer the disturb factor of the individual feature and the sunshine.4) Facial expression classificationBecause of the property of multi-mode, it's difficult to classify the facial expression. In this paper, we chose the BP neural network classification with the non-linear activation function to classify the facial expression. The network conclude three layers: input layer,hidden layer and output layer. There are eight nerve cells in the hidden layer. We chose 58 images from JAFFE database as the training swatch, and rest 60 images to test. The result of experiment show that the average classification rate of"Happy","Surprise","Sad"and"Neutral"can reach to 86.67%. In order to strengthen the reliability of the algorithm in this paper, we chose 48 images from the surrounding to test. The result of the experiment display that facial expression rate of"Happy","Surprise"and"Sad"can reach 80.56%, the performance is very well.