WiSSFormer: A Wavelet-Infused Spatial-Spectral Transformer for Superior Image Denoising

This research introduces WiSSFormer, a novel wavelet-infused spatial-spectral Transformer for image denoising that addresses the limitations of existing approaches operating exclusively in the spatial domain. WiSSFormer employs a symmetric encoder-decoder architecture incorporating two key innovations: SDC-MHA (Shift with Depthwise-aware Convolutional Multi-Head Attention) that utilizes cosine similarity for structural alignment, and SSPF-FFN (Spatial-Spectral Parallel Fusion Feed-Forward Network) that performs concurrent feature projection in both spatial and wavelet domains. The proposed cross-domain attention gate adaptively integrates complementary information across domains, enhancing detail preservation while effectively suppressing noise. WiSSFormer also addresses the high computational complexity typical of Transformer-based models by incorporating efficient architectural designs, achieving a favorable balance between performance and resource consumption. Extensive experiments on both synthetic and real-world benchmarks demonstrate that WiSSFormer consistently outperforms state-of-the-art denoising methods. Furthermore, applicationlevel evaluations, such as autonomous driving scenarios, highlight its superior generalization and practical utility in computationally constrained environments.