Single Stage Adaptive Multi-Attention Network for Image Restoration.

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• Attention-based networks are effective for image restoration but often computationally expensive with limited receptive fields.
• Existing models struggle with spatial-contextual resilience and pixel-to-pixel correspondence, hindering feature representation quality.

Purpose of the Study:

• To propose a novel, computationally efficient architecture, the Single Stage Adaptive Multi-Attention Network (SSAMAN), for image restoration tasks.
• To address limitations of existing methods by improving receptive fields and spatial-contextual robustness.

Main Methods:

• Developed the Single Stage Adaptive Multi-Attention Network (SSAMAN) architecture.
• Introduced an Adaptive Multi-Attention Module (AMAM) integrating Adaptive Pixel Attention Branch (APAB) and Adaptive Channel Attention Branch (ACAB).
• Conducted extensive experiments and ablation studies on image denoising and deblurring benchmarks.

Main Results:

• SSAMAN achieved state-of-the-art results on image denoising and deblurring tasks.
• On the SIDD dataset, SSAMAN reached 40.08 dB PSNR, outperforming Restormer by 0.06 dB PSNR with 41.02% less computation.
• Achieved 40.05 dB PSNR on the DND dataset for denoising and 33.53 dB PSNR on the GoPro dataset for deblurring.

Conclusions:

• SSAMAN offers a computationally efficient and robust solution for image restoration tasks.
• The proposed Adaptive Multi-Attention Module effectively integrates channel and pixel-wise information for enhanced feature representation.
• SSAMAN demonstrates superior performance and efficiency compared to existing state-of-the-art methods.