Research On Face Recognition Under Different Illumination Based On Deep Learning

Face recognition,as a key application field of artificial intelligence and machine vision,has high practical application value due to its non-contact,operability and simplicity,and is widely used in security,finance,entertainment and other fields.However,in real scenarios,different illuminatiaon changes still have a huge impact on face recognition,leading to a reduction in recognition accuracy,which is a problem that needs to be solved in the field.This thesis is based on deep learning algorithms to optimize and improve recognition accuracy from the follow three aspects:data set,image processing and recognition network.The main work of this thesis is as follows:(1)In order to solve the problems such as low purity,insufficient quantity,and lack of illumination conditions in the current public face datasets,this thesis designs,collects and creates a face dataset called ILL-Dataset.This dataset covers illumination colors,angles,and distances and others changes.After the collection is completed,face detection is performed using the MTCN-N algorithm to obtain the normalized face data set with a uniform size.(2)An image processing network Multi-WESPE with multi-channel information fusion is proposed to perform comprehensive image processing on images captured by smartphones using deep learning methods.Multi-WESPE uses multi-channel input and then splicing to fuse image information in the discrimination network.It realizes cross-channel information integration while retaining more original image information.The network uses multiple loss functions to combine the texture,color,content,noise and other aspects of the image to achieve overall enhancement and eventually improve the quality of the image.(3)An SK-GooLeNet recognition network with a dynamic adjustment mechanism is proposed.It adaptively adjusts the size of the receptive field according to the size of the image without manual design.The neurons in the improved network can capture the features of target objects of different scales.Secondly,the optimization activation function realizes the iterative update of parameters in the negative value interval.And we use the double loss function to accurately identify,effectively improving the model performance.