Lightweight Image Super-Resolution Reconstruction Guided by Gated Feedback and LatticeFormer

Despite the excellent performance of Transformers in image super-resolution reconstruction, their high complexity and single forward-feed information flow limit the effective mining of deep information. To address this issue, we propose a lightweight image super-resolution reconstruction model guided by gated feedback and LatticeFormer. Firstly, the reconstruction effect is significantly improved without adding extra parameters by proposing a gated feedback-LatticeFormer block as the basic block in the feedback block. Then, a gating block is designed to integrate deep global information from the previous iteration to enhance the expression of shallow features. Finally, continuous LatticeFormer blocks are constructed to improve the network’s feature extraction capability further using an attention mechanism, thus achieving effective acquisition of global features at different levels. Compared with other existing lightweight image super-resolution reconstruction methods, our method on datasets such as Set5, Set14, BSD100, Urban100, and Manga109 for <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$2 \times $ </tex-math></inline-formula>, <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$3 \times $ </tex-math></inline-formula>, and <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$4 \times $ </tex-math></inline-formula> tasks reduces the model parameters by 3%, 5%, and 7% respectively, while increasing the Peak Signal-to-Noise Ratio (PSNR) by 0.18-1.05 dB, 0.17-0.82 dB, and 0.18-1.00 dB respectively. Additionally, in the <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$2 \times $ </tex-math></inline-formula> and <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$4 \times $ </tex-math></inline-formula> tasks of the Thangka dataset, the PSNR is increased by 0.89 dB and 0.60 dB respectively. Experimental results show that our method can generate high-quality reconstructed images with complete structures and clear texture details.