Face Recognition Research Based On Curvelet Multi-orientation Feature Fusion

Face recognition is a type of biometric recognition and has become one of the most popular research topics in computer vision,pattern recognition,and biometric in recent years.Compared with other biological features,face recognition has the advantages of naturalness and non-invasion.With the rapid development of computer technology,face recognition technology has been widely studied and applied,for example,when entering the railway station,face detection,face payment,work attendance and beauty camera,etc.But the current face recognition technology still has some defects,and the recognition performance will be greatly affected in non-ideal situations such as changing lighting,facial expressions,posture,and occlusion.Wavelet transform occupies a place in many face recognition algorithms.It not only has multiresolution trait,but also has good localized analysis performance in time and frequency domain,so it is widely used in face recognition.Curvelet transform not only has the multi-resolution characteristics and time-frequency local characteristics of traditional wavelet transform,but also can display the face image more sparsely and gather the energy of the signal.Compared with wavelet transform,Curvelet transform with strong directional and anisotropic can better represent the singular features of straight lines and curves of the image.In order to obtain more effective face image features and improve the recognition effect,this paper proposes two face image feature extraction methods in the Curvelet domain,and then applies them to face recognition algorithms.The main research results of this paper are as follows:First: An adaptive weighted Curvelet gradient direction histogram face recognition algorithm is proposed.First,the face image is obtained by multiscale and multidirectional Curvelet transform coefficients through discrete Curvelet transform based on wrapping,and then the features of different directions at the same scale are encoded and fused according to the encoding method to obtain the fused amplitude domain map.The HOG(gradient direction histogram)operator is combined with the block method to obtain the histogram characteristics of the fusion image after the Curvelet transform,and the weight of each scale is calculated according to the contribution of each scale to the face recognition rate.,And then combine the weight coefficient with the HOG features of each scale.Finally,the nearest neighbor classifier(KNN)is used to classify the extracted features.Four face libraries of ORL,YALE,AR and CAS-PEAL were selected for the experiment.And the final data expresses that the improved method has a better recognition effect under the interference of partial occlusion,posture,expression,lighting changes and noise of face images.Second: A face recognition method based on Curvelet transform principal direction mode is proposed.First,the Curvelet transform of the face image is used to obtain the multi-scale and multi-direction amplitude domain map,and then the average amplitude fusion of the second and third scale amplitude domain maps is performed.Then,the asymmetric main direction pattern coding is extracted from the fused image to extract features.Then,the obtained coding features are divided into blocks and statistical histograms.Finally,the nearest neighbor classifier is used for classification and recognition.Three face libraries of ORL,AR and CAS-PEAL are selected for experiment.The experimental results show that the proposed algorithm has good recognition effect and anti-noise performance.