Abstract:  In recent years, convolutional neural network has shown good results in single image super-resolution reconstruction task, and has become the most widely used algorithm in this field. However, this algorithm fails to effectively weaken the one-to-many ill-conditioned problem and reduce the solution space range of the reconstructed image, so the effect of this algorithm in improving the image reconstruction quality is becoming more and more limited. At present, it faces the bottleneck problem, and it is difficult to improve the performance greatly. In order to effectively reduce the solution space of reconstructed images and improve the performance of reconstructed images, this paper propose an image super-resolution reconstruction algorithm based on invertible neural network. Through model design, the image degradation and reconstruction process is designed as a reversible transformation process, which effectively constrains the image solution space. The application of invertible convolution structure makes the algorithm obtain the most suitable channel arrangement rules, thus effectively improving the model performance. Experimental results on mainstream data sets show that the proposed algorithm greatly improves the accuracy of image reconstruction compared with the existing SISR algorithm, and achieves the best PSNR and SSIM performance.

Key words: super-resolution reconstruction, countercurrent model, pixel rearrangement, invertible coupling