A multi-scale enhanced large-kernel attention transformer network for lightweight image super-resolution

To address the issues of detail loss, natural texture distortion, small receptive field, overly smooth reconstructed images, and the limitations of single-scale convolution kernels in existing super-resolution models, we have designed a lightweight super-resolution network based on a multi-scale enhanced large-kernel attention transformer (MELTN). First, we propose a multi-scale shallow feature extraction module to expand the feature space, using blueprint separable convolutions instead of standard convolutions to reduce computational and memory requirements. Second, we construct the MELKAFormer module, based on multi-scale enhanced large-kernel attention and an efficient feedforward network, as the core component for efficient feature extraction. Third, we use edge-oriented convolution block and introduce a residual structure at the network output to effectively capture image edge details and texture features. Experimental results show that the proposed model maintains high image quality with significantly fewer parameters and FLOPs compared to other deep neural network models across various benchmark datasets, demonstrating the effectiveness of the model in terms of image quality and computational efficiency.