D²former: Dual-branch differential transformer for infrared image denoising.

Abstract

Infrared imaging is essential for environmental monitoring and astronomical remote sensing. However, its utility is often diminished due to sensor limitations and environmental factors that introduce noise into the images, leading to low-contrast visuals and a decreased signal-to-noise ratio. To address these challenges, we propose an innovative denoising technique inspired by the noise cancellation concept used in headphones. Our approach, named D²former, significantly enhances the quality of infrared images by adaptively applying differential mechanisms into spatial and channel attention computations. We present two primary modules: the Differential Spatial-wise Transformer Block (DSTB) and the Differential Channel-wise Transformer Block (DCTB). By emulating how noise-canceling headphones reduce ambient noise through differential signal processing, these modules create sparser and cleaner attention maps by subtracting parallel softmax responses, enabling the model to focus more accurately on critical structural elements. Building on this foundation, D²former employs a dual-branch global-local encoder-decoder architecture, where the DSTB is used in the local branch for detailed feature extraction and the DCTB in the global branch for contextual interaction. An adaptive bidirectional interaction module further enhances multi-scale feature exchange, ensuring complementary integration of local and global information. Extensive experiments on multiple infrared datasets under simulated noise conditions demonstrate that D²former consistently outperforms state-of-the-art denoising methods in terms of both visual quality and structural fidelity. The source code is publicly available at https://github.com/MingLicxy/D2former.