Research On Face Recognition Algorithm Based On HOG And Gabor Features

The recognition of face images is essentially the processing of the features of face images,while the unprocessed original face image features often have many deficiencies,such as features with more redundant information,higher feature dimensions,and features for each The ability to distinguish interference factors is poor.How to extract the key features of the face with strong anti-interference ability and low dimensionality is the most important part of the face recognition process.The criticality and completeness of the features of the image also directly affect the quality of the image recognition.In this paper,from the perspective of image feature extraction and image classification and recognition,this paper deeply studies the extraction methods,principles and application fields of several image features such as HOG features and Gabor features,and deepens the classification principle and optimization method of artificial neural networks.On the basis of comprehensive analysis,the existing algorithms and BP neural network classification algorithms are used to improve or optimize the algorithms.The main research contents and results of image recognition in this paper are as follows:Aiming at different factors such as different information entropy in different parts of the face image and different influence on recognition,an information entropy weighted HOG feature extraction algorithm is proposed.On the basis of the HOG feature extraction algorithm,the face image is firstly segmented.The purpose is to distinguish the important part and the non-important part of the face from the HOG feature extraction,and then calculate the image contained in each block.The information entropy is added as a weight coefficient to each block to form a new HOG feature.Finally,the weighted HOG feature is subjected to PCA dimensionality reduction to remove redundant information.The algorithm takes full account of the influence of different parts of the face on the recognitioneffect,and emphasizes the role of the parts with strong information entropy in the classification recognition.In the process of recognizing a face image,sometimes only extracting a single feature of a certain aspect of the face image does not achieve a better recognition effect.Aiming at this problem,an extraction algorithm based on block Gabor feature combined with HOG feature is proposed.On the basis of in-depth study of Gabor feature extraction and HOG feature extraction method,the Gabor feature extraction method is used to obtain the Gabor features of different scales and directions,and then the Gabor features are scaled according to the scale and direction.The fusion is performed,and then the merged Gabor features are segmented.Finally,the image features of the segments are further subjected to HOG feature extraction and PCA dimensionality reduction to obtain new H-G features.By introducing the idea of feature fusion,multi-level and multi-angle fusion of features is carried out.On the one hand,the integrity of face image is fully preserved,and the interference of fuzzy parts on the recognition effect is reduced.On the other hand,it increases the anti-jamming ability of the feature in the process of recognition.Aiming at the shortcomings and permissions found by traditional BP neural network algorithm in face image classification and recognition,this paper proposes additional momentum combined with gradient descent BP neural network optimization algorithm.The optimized algorithm starts from the two angles of BP neural network weight and direction.Firstly,the momentum quantity is added to each weight by additional momentum,and then the weight direction is adjusted by the elastic gradient descent algorithm.On the one hand,the combination of the two optimization methods is to avoid the problem of too long training time of neural network caused by single additional momentum optimization.On the other hand,it also avoids the problem that the single gradient descent optimization algorithm can easily make the neural network fall into the local optimal solution.The recognition efficiency of the algorithm is enhanced.