Image Blind Deblurring Algorithm Based On Deep Convolutional Neural Network

Images are one of the most important carriers of information and are used in many fields such as scientific research,medical research and criminal investigation.However,due to many reasons such as camera shake,object movement or out-of-focus disposal,the image is easy to cause blur,resulting in difficult to obtain effective information from the image.Therefore,the study of image deblurring is of great significance in scientific research and practical application.In order to improve the defuzzification performance of the computer,the blind image deblurring problem is investigated through the construction of network modules and design of deep convolutional neural network models in this thesis.The main research work is outlined below.To address the problems that the current convolutional neural network-based image deblurring methods are prone to loss of image texture details and do not distinguish feature information when processing different spatial and channel feature information,a multi-local residual connectivity attention network is constructed for image deblurring,combining residual connectivity with an attention mechanism,making the network more flexible in processing different types of information,helping to remove blur and extract Contextual information is extracted.The experimental results show that the PSNR and SSIM on the Go Pro dataset reach 31.83 d B and 0.9275,respectively.The experimental results show that the algorithm can effectively recover the texture details of blurred images and the overall image recovery is good.In order to solve the problems of loss of image detail information and large network models in the image deblurring process,a residual connected multi-branch parallel dilation convolutional network is proposed for image deblurring.By expanding the network width and using dilation convolution with different dilation factors,different features extracted from a larger perceptual field can be obtained while reducing the parameters,which helps capture more high-frequency information and reduce the model size.The experimental results show that the proposed algorithm achieves better dynamic scene deblurring performance qualitatively and quantitatively compared to the comparison algorithms.To enhance the effect of image deblurring,a two-branch feature extraction and cyclic refinement network is designed to remove image blurring in an end-to-end manner.By using different feature extraction sub-networks to extract contour features and detail features respectively,the image feature information is enriched,and the feature map is re-refined by alternately fusing contour features and detail features several times to fully fuse contour features and detail features with complementary information,which helps to obtain a clearer recovered image.The experiments show that the method in this paper achieves a better deblurring effect on dynamic scene images.