Research And Implementation On Feature Extraction Of Facial Expression Images

Facial expression can be extracted from images or videos by compute.According to the current understanding of human experience and ways of thinking,facial expression recognition can be implemented to expression classification and expression extraction,then which can be analyze human emotions from facial information.The accuracy and effectiveness of facial expression extraction is the key to whether the expression is correctly identified.In this paper,we focus on the feature extraction of face images.The research work in this paper has the following aspects:Firstly,a function module of facial expression recognition system is introduced in detail.This paper studies the principle and algorithm of image acquisition module and preprocessing module,then process some small sample collection experiments,including the following four aspects: face detection,image gray scale,image normalization,and image lighting compensation.Secondly,three common expression extraction algorithms are studied,including: Local Binary Pattern,Local Phase Quantization,and Rotation Invariant Local Phase Quantization.We use above three algorithms in the JAFFE image library to do some image feature extraction experiments,then compare extracted feature vectors and quantization histogram by experiments.In this paper,based on the RILPQ algorithm,we propose a new feature extraction algorithm by introducing directional derivative for two-dimensional Gauss kernel,which is called fusion of Gaussian derivatives RILPQ algorithm.Thirdly,SVM theory is studied and we use SVM pattern recognition software LIBSVM to complete expression classification and recognition.In this part,some programs and experiments are design by above algorithm,three parameters are studied,including filter direction,filter scale,scale radius.In order to find a group of best experimental parameters,we do a lot of experiments and achieve the expression recognition rate up to 92.57%.At the same time,in order to verify efficiency of the algorithm we proposed,two indicators are plotted including run time and recognition rate.The experimental results show that the proposed algorithm has the highest complexity and recognition rate,though the algorithm has a long run time,yet it can achieve a better effect of facial expression recognition.Finally,an algorithm for extracting expression features of motion blur face pictures is proposed based on fusion of directional derivatives and RILPQ theory.The feature extraction is experimented in the JAFFE image library after horizontal motion blur.The experimental results show that: under the condition of fuzzy length 5 pixels and scale radius 9 pixels,expression recognition rate can reach 66.10%,which is better than RILPQ algorithm recognition rate by 1.4 percentage points.