Generative Learning Approach For Face Blur Image Recovery

When cameras take face pictures,they often get blurred images due to shake,out-of-focus,etc.And when the face image is blurred,it has a great impact on advanced vision tasks such as face recognition.The purpose of face image deblurring is to recover a clear image from a blurred input image to improve recognition accuracy,etc.Therefore,the image deblurring problem is the focus of research inside the image restoration field,and the general deblurring methods do not work well on face images.In this thesis,a generative learning approach combined with a feature correction module is utilized to achieve high accuracy recovery of blurred face images,with the following main work.(1)Research on fuzzy information correction technology based on efficient self-attentionThe extraction and selection of image features is an important part of the image processing process.Fuzzy image features have different degrees of distortion and are difficult to acquire accurately.In this thesis,the Self-Attention Module(SAM)is utilized to capture the internal correlation of features in order to achieve effective correction of blurred image features and facilitate subsequent image recovery.Meanwhile,in view of the high computational overhead of the self-attentive mechanism,this thesis adopts an efficient self-attentive mechanism that achieves a balance between performance and computational overhead.(2)Generative network design for image restorationIn order to enhance the detail information of the recovered image,this thesis designs a feature fusion network(FFN)equipped with correction module for image deblurring by combining a multi-scale feature pyramid fusion approach and a progressive generation approach.The network performs layer-by-layer correction and fusion of face features,resulting in a highquality recovered image rich in detail information.In this thesis,comparison experiments were conducted with the current mainstream image deblurring algorithms to achieve SOTA results for the face recovery direction.In terms of subjective evaluation,the recovered images in this thesis are rich in details and have fewer blurred wind marks,and the five senses information is clearer and more three-dimensional,which improves the overall image recovery quality.In terms of objective metrics,this thesis compares the PSNR and SSIM metrics with other methods,where the PSNR metric exceeds the current performance optimal model by nearly 1.5 db and the SSIM metric by 1.5%.(3)Comparative experiments and ablation verification of image recovery networksThis thesis conducts comparative experiments with nearly a dozen algorithms more commonly used in the field of deblurring as well as the ablation verification of its own key modules.In terms of comparison experiments,this thesis compares four aspects: synthetic images,real images,face recognition accuracy and model running efficiency.First,in this thesis,two synthetic image test sets after face alignment and six self-made synthetic test sets are used for deblurring comparison,and the best recovery effect and the highest recovery index are achieved.Then,this thesis achieves higher recovery quality and accurate recovery of character features(e.g.,acne moles)in a recognized real test set.Finally,this thesis has the highest recovery face recognition accuracy among all deblurring algorithms.At the same time,the paper achieves high operational efficiency by using a medium-sized model.