Multi-band Video Deblurring via Efficient Chunked Additive Attention and Dynamic Channel Adaptive Module

Existing deep learning-based deblurring methods perform well in the visible spectrum but suffer from high computational costs due to complex network designs. For infrared images, texture deficiency and edge degradation hinder effective feature learning. To overcome these problems, a Dynamic Channel Adaptive Network (EDCANet) is proposed, which is based on the Dynamic Channel Adaptive Module (DCAM) and Efficient Chunked Additive Attention (ECAA) integrated into a multi—scale network with an encoder—decoder architecture. The DCAM adaptively recalibrates input frame contributions through dynamic channel weighting. Additionally, the ECAA mechanism that decomposes attention operations into partitioned spatial and frequency domains is adopted to enhance contour recovery performance. Images across multiple spectral bands are used in experiments to verify the generalization ability of the proposed method and the effectiveness of structural contour and fine-grained detail recovery. Experimental results show that the EDCANet can process images in real time at a speed of 100 frames per second, with the Peak Signal-to-Noise Ratio (PSNR) exceeding 31 dB and the Structural Similarity Index (SSIM) surpassing 0.92 on infrared datasets. For visible light datasets, its PSNR reaches above 32 dB and SSIM exceeds 0.93.