Lightweight super-resolution networks with global and local residual characteristics

Abstract

At present, there is a problem of complexity and calculation of the image super-resolution algorithm network. To improve this problem, we propose a lightweight super-resolution network that blends global and local features. First, shallow image features are extracted using a convolutional block, and secondly, deep features are extracted by multiple cascading residual feature distillation blocks GLRB, where in order to achieve a good trade-off between model performance and network parameter quantity, local features are learned by enhancing the feature selection module ESA and the balanced dual-attention module to improve model performance. Then, the extracted residual features are fused, and the reconstructed image is obtained by sub-pixel convolution sampling. The experimental results under multiple standard test data sets show that the reconstructed image performance PSNR is improved by 0.25 dB to 32.23, and the number of model parameters is 470.21 K. Compared with DRCN, CARN, IMDN, RFDN and other algorithms, the proposed algorithm has better model lightweight and image reconstruction quality.