Research On Image Restoration Method Based On Multi-scale Feature Fusion And Gated Convolution

Image restoration is an important element of research in the field of computer vision and has a wide range of applications in the fields of image editing,film and television special effects production,virtual reality,and cultural heritage preservation.The main goal of deep learning-based image restoration is to reconstruct and repair the contents of the broken region by using the a priori information of the input image.At the same time,by designing the architecture of the network model and selecting the appropriate loss function,the repaired image can be made coherent in structure and clear in detail.At present,the mainstream methods of image restoration are mainly based on generative adversarial network models and use game mechanisms to repair broken images at the semantic level.This thesis focuses on the problems of image restoration model such as loss of large-scale detail information,insufficient utilization of feature information,inability to correctly distinguish the effective pixels in the image to be restored and insufficient attention to the key regions of the image.Two image restoration methods are proposed by multi-scale feature fusion,gated convolution and multi-level attention under the generative adversarial network model.The main research and contributions of this thesis are as follows.(i)In this paper,we study three mainstream representative image restoration algorithms and experimentally compare the restoration results of the three methods and analyze the advantages and disadvantages of the three methods,and then find that the image restoration results of the mainstream representative methods have problems such as blurred details,distorted structure,restoration errors and insufficient utilization of feature information.(ii)The current representative methods all suffer from the failure to fully utilize the detail information in the image to be repaired at large scales,and rely only on the structural features after feature extraction from the coding network for repair.At the same time,the decoding network also suffers from the problem of inadequate utilization of information around the broken region.Therefore,this thesis proposes a multi-scale feature fusion image restoration method,which increases the perceptual field of the convolutional filter by adding dilated convolution to the coding network and aggregates the feature information contained in different scales of the coding network into the decoding network of the corresponding scale by jumping connection,so as to solve the problem of insufficient utilization of the information around the broken area by the decoding network in the restoration stage.Experiments prove that the multi-scale feature fusion image restoration method proposed in this thesis has good detail reconstruction capability and generates more realistic restoration results.(iii)In this thesis,we propose an attentional image restoration method based on gated convolution and multilevel spatial attention mechanism(MSAM)to ensure that the image is restored with effective pixel information and the attention to important feature regions is not strong enough.Based on this,this thesis proposes an attentional image restoration method based on gated convolution,which introduces gated convolution and Multilevel Spatial Attention Mechanism(MSAM)in a two-stage network architecture to ensure that the network model can generate restoration results with correct structure and realistic details.The experimental demonstrate that the proposed gated convolution-based attentional image restoration method recovers a better and more coherent structure and clearer details of the damaged area.