An Advanced BP Neural NetworkFace Recognition

Face recognition, beginning in the late1960s, is an important area of computerpattern recognition, image processing、analysis and understanding, computer visionand artificial intelligence, it’s one of the key technologies of biological characteristicsidentification, which develops rapidly in the nearly40years. Face recognitiontechnology has been widely used in security field of ordinary people’s living and work,such as checking on work attendance, access control, forensic and detection, it hasachieved satisfactory results, which are built conducted under ideal condition, a largenumber of experimental tests and practical experience tells us that face recognitiontechnology is far from mature under non-ideal conditions. How to eliminate externalfactors interference of illumination, facial expression, gesture and other masks, etc. toextract representative features for effective classification from high-dimensional faceis still the focus and difficulty of face recognition research at the moment. In theprocess of face recognition, the key is to image features extraction and dimensionality,principal component analysis is an effective method, but it is a statistical method, allthe pixels in the image are given the same status, including the external factors.Generally speaking, the external factors are easy to produce noise in the image edge,Curvelet transform has good skills of image sparse expression, good for de-noisingeffect, however, there is redundancy can not be ignored in scale factor. In order toextract stable and effective facial feature, we use a method combining Curvelet and2DPCA, firstly Curvelet transform the face image, we can get the low-frequencycoefficients, then2DPCA them to extract face features.With the rapid development of artificial intelligence, the artificial neural network,which has characteristics of highly nonlinear and parallelism, stronger adaptive,self-learning function and good fault tolerance and associative memory function, getsmuch attention of researchers, especially for the BP neural network, because of itssimple and actionable hierarchy, which is easy to program, furthermore, BP is widelyused and studied for its high calculation accuracy, The network uses steepest gradientdescent method and the nodes between the layers relate to each others, once thestructure is complex, the network will has slow convergence rate and be easy to fallinto local minimum points in process of learning and training, which will causeincomplete learning problem. This paper presents a face recognition method of BP neural network combining Curvelet and2DPCA, the face projection spatialdimensions of2DPCA will be selected according to BP recognition rate, the BPnetwork will be learning the most representative features in this way, the method notonly can effectively avoid “overfitting” problem to some extent, but optimize the BPnetwork structure, it will be simulated on matlab application program, the experimentuses ORL and YALE face image database, the experiment data shows that the methodmentioned in the paper has higher right rate of face recognition and robustness, it hasgood generalization ability, although it need more time to learn, its generalperformance has been be improved.