Feature Of Features:A Study On Face Recognition Based On Local Feature-inputted CNN

Deep learning has been considered as one of the most important breakthroughs in the field of artificial intelligence over the past decade.And convolutional neural network(convnet),a deep learning model,has achieved great progress in face recognition.Even so,convnet-based face recognition is still far from perfect.To the best of our knowledge,since convnet was considered as an effective feature extraction method,previous researches mostly trained convnet models using large scale of raw data,which could retain more information of input data than handcrafted features.However,this strategy encountered a bottleneck when only a small number of training samples per subject were available.This paper argues that,compared with using the raw data,training on handcrafted features,such as local binary pattern(LBP)or local derivative pattern(LDP),could be a more appropriate strategy for the convnet models,especially in the case of limited number of training samples per subject.Experimental results show that the convnet models training on handcrafted features consistently performs better than those training on raw data,even achieved several times of accuracy improvement on one-shot(with only one single training sample)experiments.The contributions of this paper were summarized as follows.1)We extended the LDP theory to meet the needs of multi-scale or multi-angle feature description.2)A LeNet-liked convnet for face recognition was proposed,which was improved from LeNet-5 convnet.LeNet-liked convnet simplified the network construct in order to accelerate training process.Experimental result showed that LeNet-liked convnet was of good performance.In order to verify the impact of different types of input for convnet,multi experiments were designed to analyze the differences of convnet,which were summarized as follows:1)By comparing local feature-inputted and raw data-inputted convnet on different face databases.We got that LBP feature-inputted convnet was of better performance.While doing one-shot experiments,LBP feature-inputted convnet achieved accuracy of 36.756%,much better than raw data-inputted convnet of 4.125%.2)There were 4 aspects,including attribute of input feature,feature of the middle layer of convnet,converges and overfitting,were considered to analyze the impact on convnet of different types of input data.Considering the experimental results on different face databases and the analysis of convnet structure,we believe that learning from feature domain maybe a better option.