Motion Blur Super-Resolution Reconstruction Based on Accelerated Deep Network

In order to meet the demand of super-resolution reconstruction of low-resolution motion blur image, an accelerated deep network architecture was proposed. In this method, the high frequency residual information between low-resolution motion blur image and high-resolution image was trained by deep network. In addition, a shrinkage layer was added in front of the nonlinear mapping structure of the series of multi-layer small-size filters to reduce dimension of the extracted feature map. Then an extension layer was added to increase dimension after nonlinear mapping. Finally, the residual information obtained from network was added to the low-resolution motion blur image to obtain the reconstructed high-resolution image. The peak signal to noise ratio (PSNR) and structural similarity (SSIM) were used to evaluate the reconstruction quality. Taking set5 test dataset as an example, the reconstructed mean result is 29.05/0.8339, which is greatly improved compared with VDSR and other deep learning methods. The reconstruction speed is also improved compared with these methods. It shows that this method can effectively solve the problem of super-resolution reconstruction of motion blur image.