Image Deblurring Based On Convolutional Neural Networks

The captured images often contain undesired blur,such as,object motion blur,camera shake blur,scene depth variation,etc.The goal of image deblurring is to restore the sharp image from the blurry one which is a challenging computer low-level vision task as the highly ill-posedness.The existing technology has some limitations in image deblurring,including insufficient feature information extraction,slow speed,convolution locality and the collection of real blur dataset etc.In the past two years,with the development of deep learning,Convolutional Neural Network(CNN)methods are widely used in all aspects of image processing.This ene-to-end image processing method,on the one hand,makes some image restoration problems improved significantly,on the other hand,it improves the speed of image processing to achieve real-time processing effect.The main innovative ideas are as follows:1.A fast deblurring network based on progressive processing strategy is proposed.Progressive Deblurring Network(PDN)cascades several sub-networks and realizes image restoration step by step.The progressive processing strategy could fully utilize multi-level prior features.The deblurring task is sensitive to pixel level information,so the image feature extraction network is re-designed.Through a variety of short-cut connections to get faster processing speed,and the reduction of parameters and inter layer operations also helps to speed up and achieve real-time test results.2.Because of the blur degree of different positions in an image is inconsistent,the kernel with uniform size will process the whole image with no focus,while the larger convolution kernel will introduce more parameters.In order to balance the relationship between the above-mentioned two aspects,the image quality and processing speed,the multi-scale feature extraction network(MFN)is introduced.Meanwhile,since the convolution is too local,attention mechanism is introduced to selectively amplify valuable information and suppress some useless features.It also processes the feature map at the U-net skip connections to help image restoration.This method makes full use of the features of different scales,including the shallow and deep semantic features,and can get better deblurring effect.3.A hierarchical network based on improved attention mechanism is proposed.For the improvement of attention mechanism in the above chapter,using both local and global information to enhance the information extraction ability of the network;secondly,the input image is divided into several blocks,which is helpful to the fusion of fine and coarse features.This chapter verifies that the real dataset image has better visual effect and higher evaluation index than other methods.In general,the proposed models are tested in several datasets,and compared with the existing state-of-the-art models based on convolutional neural network.The results show that the single image deblurring models proposed in this paper can restore the blurred image better,produce higher peak signal-to-noise ratio and keep the edge well.At the same time,the human visual effect is more pleasure than others.