Image deblurring of dynamic scene based on attention residual CODEC network

Abstract

The image deblurring technology of dynamic scene is a challenging computer vision problem. The blurry images not only affect the subjective perception but also affect the performance of the subsequent intelligent analysis. An image deblurring method of dynamic scene based on attention residual CODEC network was proposed. Firstly, in the coding stage, many residual modules were used to extract the features, and the spatial attention module was added to perceive the blurry spatial position information. Then, a global-local residual connection strategy in the internet was adopted to fuse the multi-layer convolution features to reduce the information loss. Finally, a restored image with clear edges and structure was generated in the decoding stage. The experimental results show that the peak signal-to-noise ratio (PSNR) obtained on the public data set is 31.76 dB, and the structural similarity index measure (SSIM) value is 0.912. Both the objective and subjective quality evaluations indicate that the proposed method can effectively recover the clear images containing abundant edge contour information, which obtains the optimal performance in the compared algorithm.