MobileDMD-Net : Deformable MobileNet Backbone for Efficient Motion Deblurring

This paper introduces MobileDMD-Net, a lightweight deep neural network designed for real-time single-image motion deblurring in dynamic scenes. Unlike conventional approaches that rely on computationally intensive architectures or handcrafted priors, MobileDMD-Net employs a compact, encoder-decoder design optimized for efficiency without com- promising image restoration quality. The model integrates a blur-aware attention mechanism and frequency-guided feature fusion to effectively disentangle motion blur across varying spatial scales. Extensive experiments conducted on the Go- Pro and Köhler datasets demonstrate that MobileDMD-Net achieves state-of-the-art performance, with PSNR and SSIM improvements of up to 3.17 dB and 0.034, respectively, over strong baselines. Notably, the model requires only 4.10<br/> GFLOPs and achieves an inference time of 0.014 seconds, enabling deployment on real-time systems. These characteristics make it well-suited for latency-sensitive applications such as onboard image enhancement for unmanned aerial vehicles (UAVs) and edge computing environments. The proposed method represents a compelling trade-off between accuracy and computational cost, advancing the practical feasibility of high-quality motion deblurring in real-world scenarios.