Research On Face Recognition Method Based On Feature Fusion

In 2020,a serious new crown epidemic occurred at home and abroad.In order to facilitate the management of the flow of people,many stations,campuses,communities and other places with high traffic have installed face detection equipment,which indicates that face recognition technology has become identity recognition.One of the important technical means.However,in an environment with complex lighting,the accuracy of face recognition technology will drop rapidly,and the changing lighting environment is already one of the main external factors that affect the accuracy of recognition.In this thesis,the three aspects of face image preprocessing,face invariant feature extraction and feature fusion are studied and the algorithm is optimized.Finally,the goal of improving the recognition rate under complex lighting conditions is achieved.The main work of this thesis is as follows:(1)The Gabor self-quoted image algorithm model is studied and improved.Due to the insufficient selection of image features in traditional self-quoted image algorithms,the problem of high similarity of processed face images appears.This thesis first uses an improved multi-scale and multi-directional Gabor algorithm to extract features from the collected face images,and then performs de-redundancy processing on these features to reduce the repetition of the same features,and then uses the nearest neighbor classification method in Extended Yale B and CMU PIE face database for experiments.It can be concluded from the data comparison that the improved algorithm improves the accuracy of face recognition technology.(2)Research the feature extraction algorithm of face image and perform dimensionality reduction processing.Aiming at the problem that the single feature extracted cannot accurately describe the face under the condition of changing illumination.In this thesis,the MTLBP algorithm is used to extract the facial features of the face,and the second-order Laplacian algorithm(Lo G)is used to extract the details of the edge of the face,so as to optimize the feature information of the face image and make it more accurate.In the case that the total number of identifications has a small change,such as the management of personnel in and out of enterprises and universities,using the PCA algorithm to perform dimensionality reduction processing on feature extraction can achieve the effect of greatly reducing the amount of calculation.(3)The method of determining the weight of facial features extracted by different algorithms is studied.Aiming at the problem of poor application performance of fixed weights in different lighting scenarios,a method for dynamic fusion of facial features based on standard deviation is designed.This thesis uses the size of the standard deviation as the amount of information occupied by the facial features to achieve the weight setting,and dynamically fusion the feature information extracted by the MTLBP algorithm and the Lo G algorithm to reduce the interference of noise such as light.The experimental results show that the fusion of the features extracted by the two algorithms improves the accuracy of face recognition.Based on the above research,this algorithm based on feature fusion improves the accuracy of face recognition.