Super-resolution reconstruction based on improved diffusion probabilistic model

Abstract

Despite deep learning models have made significant breakthroughs in the magnification and precision of single image super-resolution (SISR) reconstruction. However, current methods which focus on extracting rich texture details, always ignores the influence of byproduct of artifacts in high factor super-resolution construction. Therefore, a super-resolution method based on improved diffusion probabilistic model (DPM) is proposed to eliminate artifacts while obtain texture details. Firstly, residual skip path (ResPath) is designed to enhance the constrain of initial information on the reconstruction result to mitigate artifacts. In addition, overfitting may also generate artifacts, so blueprint separable convolution (BSConv) is introduced to reduce redundant network parameters. Secondly, in order to retain the texture details, the efficient channel attention (ECA) block and the enhanced spatial attention (ESA) block are incorporated to extract more comprehensive channel and spatial implicit information. The effectiveness of this method is verified on 4 public datasets with ×8 factor. Compared with other advanced SISR methods, the proposed method achieves the best performance on perceptual index (PI) and visual perceptual quality.