Lightweight multi-scale distillation attention network for image super-resolution

Convolutional neural networks (CNNs) with deep structure have achieved remarkable image super-resolution (SR) performance. However, the dramatically increased model parameters and computations make them difficult to deploy on low-computing-power devices. To address this issue, a lightweight multi-scale distillation attention network (MSDAN) is proposed for image SR in this paper. Specially, we design an effective branch fusion block (EBFB) by utilizing pixel attention with different kernel sizes via distillation connection, which can extract features from different receptive fields and obtain the attention coefficients for all pixels in the feature maps. Additionally, we further propose an enhanced multi-scale spatial attention (EMSSA) by utilizing AdaptiveMaxPooland convolution kernel with different sizes to construct multiple downsampling branches, which possesses adaptive spatial information extraction ability and maintains large receptive field. Extensive experiments demonstrate the superiority of the proposed model over most state-of-the-art (SOTA) lightweight SR models. Most importantly, compared to residual feature distillation network (RFDN), the proposed model achieves 0.11 improvement of PSNR on Set14 dataset with 57.5% fewer parameters and 20.3% less computational cost at x4 upsampling factor. The code of this paper is available at https://github. com/Supereeeee/MSDAN.