MDMB-YOLO: A liquid crystal display defect detection method using Multi-Differential Fusion and multi-branch feature pyramid

To address the challenges of coexisting defects at multiple scales and the tendency for small defects to be missed in LCD defect detection, this paper proposes a novel detection algorithm. The method first designs a Multi-Differential Fusion Module (MDFM), which enhances sensitivity to small defects (especially dot defects) by integrating multiple differential sensing strategies. Second, a multi-branch fusion efficient feature pyramid network (MFEFPN) is constructed. Leveraging a multi-branch structure and efficient fusion mechanisms, this network effectively mitigates information loss and feature interference issues inherent in traditional feature pyramid networks. To further balance accuracy and computational efficiency, we designed an Adaptive Shared Lightweight Detection Head (ASLD), which maintains excellent detection accuracy while significantly reducing the number of parameters and computational complexity (GFLOPs) through a parameter-sharing mechanism. Additionally, geometric constraint terms are incorporated into the loss function to further enhance the localization capability of defect boundaries. Experimental results show that the proposed MDMB-YOLO achieves an accuracy of 85.2%, with a 4.4% improvement in accuracy, a 3.3% improvement in recall rate, a 2.8% improvement in mAP50, and a 0.9% improvement in mAP50-95 compared to the baseline model. The number of parameters and GFLOPs were reduced by 23.3% and 8%, respectively, compared to the baseline model, indicating that this approach offers both accuracy and efficiency advantages in LCD defect detection tasks. The dataset used in this study has been publicly released, and we encourage its use for related research in accordance with the platform’s terms at: https://aistudio.baidu.com/dataset/detail/358247/settings .