Gabor Wavelet Networks-Based Face Recognition Research

Face recognition is one of the popular tasks in some fields such as pattern recognition, image processing, computer vision, neural networks and cognitive science in recent years. It can be widely applied in records management systems, security verification systems, credit card verification, criminal identity recognition, monitoring in bank and customhouse, human-computer interaction, etc.This paper did some theoretical and experimental research work on facial feature extraction and recognition algorithms. The experimental results showed that the method we proposed, which introduces pupils' locations information into the process of face feature extraction implemented by Gabor wavelet networks, did improve the efficiency of feature extraction. The research work in this paper mainly includes the following several respects:Firstly, we studied the method of elastic bunch graph matching and did some experimental and analytical work. According to the principle of elastic bunch graph matching, we did two experiments on issues of feature point localization and if the same feature point still has highly similar Gabor feature vector when the facial expression or head angle has changed. We analyzed the experimental results and pointed out the shortcomings of elastic bunch graph matching method and some improvement ideas. The effective feature representation ability of Gabor wavelet networks can improve the abundance problem of elastic bunch graph matching in feature extraction.Secondly, we did some research and experiment work on eyes detection. We used gray-scale mathematical morphology and Hough transform to detect eyes and position pupils. The flow chart and some results of eyes detection experiment are showed.Thirdly, we proposed introducing pupils' locations information into the process of face feature extraction implemented by Gabor wavelet networks in order to improve the efficiency of feature extraction. Gabor wavelet networks (GWN) are combining the advantages of RBF networks with the advantages of Gabor wavelets: GWNs represent an object as a linear combination of Gabor wavelets where the parameters of each of the Gabor functions (such as orientation and position and scale) are optimized to reflect the particular local image structure. The pupils' locations information we proposed to introduce was used in two ways. Firstly, it was used to form the T-shape initial location distribution of Gabor wavelets in network optimization phase for the purpose of extracting more useful feature for recognition given a certain amount of wavelets. Secondly, it was used as locating information in the reparameterization phase to greatly simplify the reparameterization procedure. After feature extraction by Gabor wavelet networks, this paper adopted the method of kernel associative memory (KAM) to classify the feature. The experimental results showed that utilizing pupils' locations information could obviously improve the efficiency of feature extraction, and that compared to methods of Euclidean Distance, normalized cross correlation and nearest feature line (NFL), KAM achieved a better recognition rate.