A Dual-Branch Spatial Interaction and Multi-Scale Separable Aggregation Driven Hybrid Network for Infrared Image...

Area of Science:

• Computer Vision
• Image Processing
• Deep Learning

Background:

• Single Image Super-Resolution (SISR) aims to enhance image resolution.
• CNNs and Transformers excel in visible image SR but face challenges with infrared images due to noise and limited long-range dependency modeling.
• Infrared images present unique challenges like low signal-to-noise ratio and blurred edges.

Purpose of the Study:

• To develop an effective hybrid neural network for infrared image super-resolution reconstruction.
• To address the limitations of existing CNN and Transformer models in infrared imaging.
• To improve detail sharpness and visual quality of infrared images.

Main Methods:

• Proposed RDSR (Residual Dual-branch Separable Super-Resolution Network), a hybrid architecture.
• Integrated multi-scale depthwise separable convolutions with shifted-window self-attention.
• Introduced Dual-Branch Spatial Interaction (BDSI) and Multi-Scale Separable Spatial Aggregation (MSSA) modules.

Main Results:

• RDSR demonstrated superior performance in terms of PSNR and SSIM for ×2 and ×4 upscaling.
• Outperformed state-of-the-art CNN-based (EDSR, RCAN, RDN) and Transformer-based (SwinIR, DAT, HAT) methods.
• Experimental validation on multiple public infrared image datasets confirmed effectiveness.

Conclusions:

• The proposed RDSR network effectively reconstructs high-resolution infrared images.
• The hybrid approach combining convolutions and self-attention is highly effective for infrared SR.
• RDSR offers a promising solution for enhancing infrared image quality and detail restoration.