Lightweight Asymmetric Convolutional Residual Network for Efficient Motion Image Deblurring

Motion image deblurring remains challenging because many existing models rely on complex architectures, leading to high computational cost and parameter redundancy, particularly under non-uniform blur in real-world scenes. To mitigate these limitations, we propose a Lightweight Asymmetric Convolutional Residual Network (LACR) for efficient motion image deblurring. LACR introduces an asymmetric convolutional residual module that combines local spatial embedding with horizontal and vertical asymmetric refinement, enabling direction-sensitive blur modeling with reduced spatial redundancy. A shallow deep feature fusion mechanism is further designed to integrate low-level convolutional cues with deep restoration representations, thereby complementing low-frequency structural information with high-frequency texture details. Experiments on four benchmark datasets show that LACR improves the reconstruction of edges, textures, and structural details while maintaining a lightweight design. Compared with representative lightweight deblurring methods under consistent evaluation settings, LACR achieves an average PSNR gain of 0.38 dB and reduces computational cost by more than 20%. Quantitative and qualitative results demonstrate that LACR achieves a favorable balance between restoration quality and computational efficiency.