DCT-SR: Single Image Super-Resolution via Deeply Coupled Transformer-Enhanced Network

Single image super-resolution (SISR) is a classic task in computer vision, aiming to reconstruct high-resolution (HR) images from low-resolution (LR) images. Deep neural networks (DNNs) perform exceptionally well on the SISR task by learning complex nonlinear mappings in an end-to-end manner. Transformer-based methods have made significant progress due to their ability to model global information, but the local receptive field of windowed self-attention limits their capacity to capture long-range dependencies, affecting the accuracy of detail reconstruction. To address these issues, this paper proposes a Deeply Coupled Transformer Enhancement Network (DCT-SR). DCT-SR extracts long-range image features through a deep auxiliary network, making full use of global context information; it designs an Adaptive Fusion Module (AFM) to dynamically calibrate the contribution weights of local features from each Transformer layer and the deep global features, achieving efficient local-global feature coupling; it introduces discrete wavelet convolution (DWT) to enhance shallow high-frequency feature representation, improving texture and edge detail reconstruction. Experiments show that DCT-SR significantly outperforms existing methods on multiple benchmark datasets, demonstrating its effectiveness and generalization ability.