Font Size: a A A

Research On Face Recognition Based On Fusion Features Of CNN And SVM

Posted on:2022-05-05Degree:MasterType:Thesis
Country:ChinaCandidate:Q Q LiangFull Text:PDF
GTID:2518306473491944Subject:Master of Applied Statistics
Abstract/Summary:PDF Full Text Request
With the development of the information age,face recognition is widely used in other fields,including identity authentication,criminal investigation,and video surveillance.Face recognition is simple and easy to implement,but it's easily affected by light,posture,etc.because of the subject characteristics of face recognition.And it is hard to collect lots of samples under non-ideal conditions,and face recognition technology with small samples is in urgent need of development.In this dissertation,the research on face recognition with small samples is carried out.Its methods primarily contain traditional recognition methods and deep learning methods nowadays.Among them,the traditional method is mainly realized by constructing different classification features and classifiers.With the rise of deep learning,deep network technology has also been used more in face recognition,which has promoted the development of face recognition.Deep learning has strong learning capabilities and can adaptively adjust the learned image features.But deep learning requires a large sample size,and has poor generalization ability for small-scale samples and there is an over-fitting phenomenon.Therefore,studying small-sample face recognition algorithms for these problems has far-reaching significance.This dissertation raises a method based on CNN and SVM for small sample data sets.Before the recognition,use the adaptive histogram equalization that limits the contrast to process the images and weaken the impact of light.To improve the robustness of the model to small samples,a small CNN is first constructed to get image features.Finally,the results are obtained by using SVM with stronger generalization ability for small samples.Combine the learning ability of the CNN with the adaptability of the SVM to small samples.Add CBAM attention mechanism while creating feature extraction network,so that the network adaptively focuses on the feature parts that have a greater impact on classification.In this way,the bad effects of facial posture changes and facial occlusion on classification are reduced.At the same time,in order to enhance detailed information such as the contour,the weighted features from different depths of CNN are combined into the final classification feature and input to SVM.Experiments were carried out on two small sample data sets with changes in pose,light and facial occlusion,so as to prove the validity of the method.First,conduct experiments to explore the influence of feature weighted fusion weights on the classification effect,and obtain the optimal weighting coefficient.Secondly,compare the classification results with and without feature fusion.The results demonstrated that the feature weighted fusion can obviously enhance the classification effect.Finally,the algorithm in this dissertation is compared with the existing algorithms,and the results demonstrated that the algorithm in this dissertation has better classification effect.
Keywords/Search Tags:Feature fusion, Small sample, Convolutional neural network, Attention mechanism, Support vector machine
PDF Full Text Request
Related items