Research On Image Inpainting With Dense Propagation And Multi-resolution Fusion Learning

With the development of science and technology,digital images are becoming more and more important in human life.However,digital images are easily affected by many natural or unnatural factors in the process of acquisition and transmission,which can lead to image damage.When humans take pictures through transparent media such as glass or windows,the images often contain reflective images,which makes the image quality seriously affected.Although image inpainting techniques have been greatly improved with the rise of deep learning techniques,the results of current methods are still inadequate for challenging tasks such as defective image restoration and reflective image inpainting.In this thesis,we deeply analyze the current methods for defective image and reflection image representative inpainting and propose restoration methods for defective images and reflection images respectively around the shortcomings of current mainstream methods.The main work of this thesis is as follows:This thesis systematically expounds the mathematical model of image inpainting.Three current defect image inpainting methods(CA,GMCNN,PEN-NET)and reflection image removal methods(BDN,Rm Net,DAD)were studied and analyzed respectively.Further analysis of the experimental results of the mainstream methods shows that the results of the defect image inpainting method still have the problems of unreasonable structure and fuzzy details.The problem of reflection image removal is information loss and reflection is still visible.In view of the shortcomings of these mainstream methods,this thesis puts forward the corresponding improvement ideas.The existing defective image inpainting methods have the problems of unreasonable structure and rigid details.The dilated convolution fails to fully perceive the contextual information of the image because the feature size of the dilated convolution input to the inpainting network is too large.Therefore,the inpainting network cannot effectively use the effective information of the non-defective area to fill the defective area.In this thesis,a defective image inpainting method based on dense propagation is proposed.In this method,the advanced features of the inpainting network are compressed and encoded into more compact features by constructing a multi-scale dense extended convolutional network.In order to make full use of the context information of the undamaged region,compact features at different scales are propagated with dense connections through extended convolution.Therefore,the structure and details of defective image inpainting results are reasonable and smooth.Experimental results show that compared with the current representative defective image inpainting methods,the proposed method based on dense connection has stronger reconstruction ability in both structure and detail.Information loss and visible reflection still exist in the current mainstream reflection image de-reflection methods.This is mainly because the de-reflection network architecture only relies on low-resolution feature recovery and jump connection of the same resolution feature to obtain high resolution representation,but fails to use the multi-resolution representation information of the image to guide the representation of high resolution representation information.In this thesis,a reflection image inpainting method based on multi-resolution fusion learning is proposed.In this method,a parallel multi-resolution subnet is constructed to repeatedly fuse the multi-resolution representation information at multiple scales.This makes the final high-resolution representation,including structure and details,more accurate,enabling inpainting of reflection images.Experimental results show that compared with the current mainstream reflected image de-reflection method,the proposed reflected image inpainting method based on multi-resolution fusion learning can reconstruct the non-reflective image with higher quality.