Research On The Occlusion Of Glasses In Face Recognition

In recent years,artificial intelligence has been a high degree of concern and rapid development,with the rapid development of artificial intelligence,face recognition has become a very important research field in computer vision.Face recognition performance tends to be influenced by the following factors:light conditions,face posture,facial expression and facial mask,etc.Glasses as one of the most common obstructions tend to have great influence on the performance of face recognition and very wide range of glasses make this problem more complicated,and how to effectively remove the glasses in the image to become a problem to be solved.Most of traditional methods adopt PCA to reconstruct the eyeglassless faces,but the PCA needs to detect glass areas before reconstruction which itself is a great challenge and there will be obviously discontinuity between the boundary of occluded and disoccluded areas.Based on the above problem,we propose an algorithm of Automatic glass removal via Skip-Connect Deconvolutional Networks.The main work and innovations of this thesis are:1.The paper analyzes the defects of traditional PCA algorithm,the whole image reconstruction based on PCA makes the error spread to the whole image and using PCA will lost high-frequency information,which leads to the lost of local details in images,and puts forward an automatic glass removal via skip-connect deconvolutional network to obtain the more natural faces without glasses.2.Glasses removal requires both the spatial information captured in lower layers and the semantic information captured in upper layers.However,general convolutional neural network merely consists of convolution and pooling structure,which easily loses spatial information captured in early layers,and deconvlutional neural network merely reverses the pooling operation,which limits the information that can be transferred.On the contrary,common skip connections architectures is effective in feature fusion.However,if there are a variety of shelters in image,common skip connections also cannot distinguish them.Based on this,we propose a kind of bottom-up/top-down process to effectively merge the spatially rich information from low-level features with the high-level semantic information.3.At last,we compare the algorithm with PCA and Deconvolutional network.Experiments verify the effectiveness of our method.We also analyze the effect of glass removal on face recognition performance.