Research On Face Flaw Detection Based On Deep Learning

With the popularity of smart phones and the development of the Internet,more and more people share daily lives through the Internet.Many young people expand their social circle by uploading selfie images to social platforms to get comments or likes from others.Most people want to show their best side to others,so there is a growing demand for face flaw detection and repair.In recent years,with the rapid development of deep learning,more and more image processing problems have begun to use technology related to deep learning.This thesis studies the target detection algorithm based on deep learning,and improves the target detection model based on the face flaw detection scene.In addition,this thesis proposes a model compression scheme for mobile applications,and implements a complete face flaw detection system.The main work and innovations of this thesis are as follows:(1)A face flaw detection model based on location correlation loss function and improved for small target detection is proposed.Firstly,based on YOLOv3,the network is improved to detect small targets of face flaw by adding a branch corresponding to higher resolution feature maps to retain more details in the original images.Subsequently,by using a loss function based on location correlation,the location information between the prediction box and the actual box is added to the training process of the network,which makes the network training converge faster and improves the accuracy of the network.Finally,this thesis creates a special dataset for face flaw detection,and retains the accuracy of the model by data augmentation and K-means clustering.The experimental results show that this thesis improves the network performance while speeding up the network training,and the F1 score and mAP are improved by 12.1%and 10.5%respectively with the same training epochs.(2)For mobile applications,a channel pruning model compression algorithm based on network architecture search is proposed.Firstly,channel pruning is considered as a special method of neural architecture search,and only the structure of the network is retained after pruning.Subsequently,based on an existing channel pruning algorithm,this thesis improves it on sparse training and pruning.The variable sparse coefficient is used to retain the accuracy of the network in the sparse training,and the protection threshold is set to solve the problem of a sharp decline in model performance when the pruning rate is too high.The experimental results show that the variable sparse coefficient reduces the limitation of loss function on network parameters,and the existence of protection threshold also retains the accuracy of the model.At the same time,the idea of neural architecture search also effectively prevents the network from falling into the local optimal value during training.The pruning algorithm in this thesis achieves a compression rate of 11.8%for the model volume while retaining the accuracy of the model.(3)A complete face flaw detection system is designed and implemented.The system is divided into three stages:image preprocessing,face flaw detection and image post-processing.Image preprocessing includes determining the face area to be detected from the original input image and image segmentation;face flaw detection mainly uses a trained model to detect the segmented images separately;image post-processing includes image stitching and result deduplication.Taking a test image as input,the performance indicators of each stage are obtained,which verifies the practicability of the system.