Research On Multi-scale Feature Fusion Image Deblurring Algorithm Based On Attention Mechanism

As an important medium for recording and transmitting information,images are widely used in the field of artificial intelligence.During the acquisition process,the images are inevitably blurred by camera shake or object motion,which is detrimental to the further utilization of the images.Therefore,how to restore a blurred image to a clear image has become a popular topic of image processing.Deep learning methods are becoming the tool of choice for image deblurring tasks due to their powerful feature representation capabilities.In this thesis,we analyze the characteristics of blurred images at different scales,use the attention mechanism to locate the blurred regions effectively,and integrate multi-scale information to improve the deblurring performance of the network.The details of the research will be expanded below:(1)An asymmetric image deblurring network based on spatial attention feature fusion is proposed.To use the information of the image itself more effectively,the network uses three feature extraction subnets,which with the same structure,to obtain information of multi-scale blurred images separately.To fuse these multi-scale blurry features,we design a feature fusion mechanism in the reconstruction subnet,which use the information from different feature extraction subnets for spatial attention inference.Under the guidance of the spatial attention,the detail information of smaller scale is added to the blurred regions that have a greater impact on the image quality,and finally the integrated features are used to reconstruct a clear image with the original resolution.The proposed method achieves good results on two public datasets,which validates the effectiveness of the model.(2)A spatial attention based multi-scale cross network.Through experimental analysis,we find that the multi-scale strategy can bring improvement in deblurring performance,but there is a large amount of redundancy,which limits the recovery efficiency.By optimizing the multi-scale network structure,we design a multi-scale crossover structure,which uses the information from different scales for spatial attention inference.And under the guidance of the attention map,the information from smaller-scale is added to the feature extraction process of larger-scale blurred image.We locate the blurred position by computing the residual map between the real clear image and the blurred input image,which helps to supervise the attentional inference process in the multi-scale cross structure.Experiments on the public dataset verify that the proposed model maintains great deblurring performance,and improves the recovery efficiency more effectively.