Transformer models have shown remarkable performance in image super-resolution. However, existing Transformer-based methods often exhibit limitations in capturing fine-grained details and incur high computational costs. To address these challenges, this paper proposes a novel lightweight super-resolution method based on Scattering Processing and Feature Interaction (SPFI). Our framework introduces two novel modules. First, a Scatter Pre-process Module (SPM) decomposes input features into high- and low-frequency components using a Dual-Tree Complex Wavelet Transform (DTCWT). These components are then processed via an Einstein Mixing strategy, which enhances fine-grained detail extraction while reducing model complexity. Second, a Cross-token Integration (CTI) block facilitates multi-scale feature fusion in a computationally efficient manner. It employs depth-wise separable convolutions with varied kernel sizes and strides to enable structured interaction among tokens. Extensive experiments demonstrate that the proposed SPFI outperforms current state-of-the-art super-resolution methods in reconstruction quality while maintaining a lower computational complexity. Additionally, comprehensive evaluations including inference speed and noise robustness further validate the practical value of SPFI.
Zheng et al. (Thu,) studied this question.