A Self-Attention Transformer-Based Architecture for Robust Image Deblurring

Abstract

Image deblurring is a challenging task in computer vision because it is a difficult and spatially variant problem. This work presents a Transformer-based architecture that utilizes self-attention mechanisms to capture useful long-range dependencies in an image. Unlike traditional convolutional approaches, the proposed model employs an encoder-decoder architecture, in which the encoder is used to obtain hierarchical multi-scale features and the decoder is used to progressively rebuild the latent sharp image with improved quality and structural fidelity. The self-attention mechanism facilitates global context modeling, leading to improved sharpness and great minimization of artifacts. Moreover, the technique does not require explicit prior knowledge of the blur kernel and is therefore robust to various real-world blur conditions. Experimental results demonstrate that the proposed approach achieves superior performance in terms of Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM), outperforming conventional and recent learning-based deblurring techniques.