D²former: Dual-branch differential transformer for infrared image denoising

Area of Science:

• Remote Sensing
• Image Processing
• Artificial Intelligence

Background:

• Infrared imaging is crucial for environmental monitoring and astronomical remote sensing.
• Image quality is often degraded by sensor noise, reducing contrast and signal-to-noise ratio.
• Existing denoising methods struggle to effectively address these challenges.

Purpose of the Study:

• To introduce D²former, a novel infrared image denoising technique.
• To enhance infrared image quality by adaptively reducing noise.
• To improve the utility of infrared imaging for critical applications.

Main Methods:

• D²former utilizes differential mechanisms within spatial and channel attention computations.
• Key modules include the Differential Spatial-wise Transformer Block (DSTB) and Differential Channel-wise Transformer Block (DCTB).
• A dual-branch global-local encoder-decoder architecture with adaptive bidirectional interaction is employed.

Main Results:

• D²former effectively creates sparser and cleaner attention maps by subtracting parallel softmax responses.
• The method accurately focuses on critical structural elements in infrared images.
• Experiments show D²former outperforms state-of-the-art methods in visual quality and structural fidelity.

Conclusions:

• D²former offers a significant advancement in infrared image denoising.
• The technique enhances the reliability of infrared imaging for environmental and astronomical applications.
• The proposed method provides superior performance in noise reduction and detail preservation.