Enhanced Deblurring for Smart Cabinets in Dynamic and Low-Light Scenarios

In this paper, we propose a novel method to address dynamic blur and low-light issues in smart cabinets, which is named the MIMO-IMF (Multi-input Multi-output U-Net Integrated Motion Framework). This method combines a Frequency-Domain Adaptive Fusion Module (FDAFM), built on the blind deblurring framework MIMO-UNet (Multi-input Multi-output U-Net), to improve the capture of high-frequency information and enhance the accuracy of blur region recovery. Additionally, a low-light luminance information extraction module (IFEM) is designed to complement the multi-scale features of the FDAFM by extracting valuable luminance information, significantly improving the efficiency of merchandise deblurring under low-light conditions. To further optimize the deblurring effect, we introduce an enhanced residual block structure and a novel loss function. The refined multi-scale residual block, combined with the FDAFM, better restores image details by refining frequency bands across different scales. The New Loss Function improves the model’s performance in low-light and dynamic blur scenarios by effectively balancing luminance and structural information. Experiments on the GOPRO dataset and the self-developed MBSI dataset show that our method outperforms the original model, achieving a PSNR improvement of 0.21 dB on the public dataset and 0.23 dB on the MBSI dataset.