A Two-Stage Coupled Learning Network for Image Deblurring

Image deblurring is a fundamental task in computer vision, aiming to reconstruct sharp images from their blurry counterparts. Most existing deblurring methods based on deep learning adopt a uniform and fixed single-residual learning framework, failing to address the intricate degradation of image content, including delicate edge contour details and color distortion (especially in the case of complex motion blur). To address these problems, we propose a Two-Stage Coupled Learning Network (TCLNet) to autonomously learn the difference and intrinsic interaction between low-frequency features (content and color information) and high-frequency features (edge and texture detail information), helping the network better understand how to reconstruct edge details and color characteristics while maintaining computational efficiency. Especially, we partition the deblurring process into two stages. In the first stage, we introduce a novel two-decoder architecture with collaborative learning to preliminarily decouple blur features and mitigate the learning complexity of the network. In the second stage, we propose a coupled learning module (CLM) and a feature enhancement block (FEB) to constrain the network to learn features from different domains. In this way, our proposed method can restore more realistic edge details and color characteristics at low computational costs. The extensive experimental results demonstrate that our approach outperforms previous state-of-the-art methods in terms of accuracy and computational efficiency.