Global-Local Adaptive Attention-Based U-Net for Single Image Deblurring

Image deblurring is a classical ill-posed problem aimed at recovering a clear image from a blurred one. The recent application of Transformers to image restoration tasks has significantly improved performance, thanks to their powerful long-range dependency modeling capabilities. However, the high computational cost of self-attention mechanisms limits their practical use. Additionally, the training cost for Transformer models is considerable. To address the issues of high computational cost and training cost, this paper proposes a novel Global Local Attention (GLA) mechanism that combines global and local adaptive attention. The mechanism uses 1D adaptive pooling operations to capture global context information, while fixed-size 1D pooling with different kernel sizes (7 and 11) is used to capture local details, effectively addressing both long-range dependencies and fine-grained information. Furthermore, a larger receptive field is crucial for effective image deblurring. To this end, we introduce a Feature Enhancement Module (FEM) with an expansive convolutional kernel size that serves as a bottleneck layer to further improve recovery quality. Experimental results demonstrate that our proposed model achieves competitive performance on several benchmark datasets, advancing the field of image restoration.