A Method for Coal Flow Foreign Object Detection and Mineral Semantic Segmentation Based on the Lightweight Multitask Model PVDNet

To address the separation of mineral semantic segmentation and foreign object detection in coal mine conveyor belt detection, and the inability of traditional single-task models to balance small target detection and complex boundary segmentation, this paper proposes a lightweight multi-task model, PVDNet, for simultaneous optimization of multi-category mineral segmentation and foreign object detection. Firstly, a coal mine semantic segmentation dataset (CsSS) comprising 6 mineral categories and 4 foreign object categories was constructed, and MobileSAM-assisted annotation and controllable synthesis were employed to mitigate sample scarcity and category imbalance. Secondly, based on PIDNet, an improved TinyViT backbone was introduced, along with a hierarchical perception multi-scale fusion framework and a dual-attention FusedBlock module, to enhance feature extraction and collaboration between semantic details. For gradient conflict in multi-task learning, a hierarchical gradient pruning strategy integrated with PCGrad was proposed to improve convergence speed and training stability; a differentiated multi-task loss function was also designed to balance optimization objectives. Experimental results show that PVDNet achieves 95.16% mIoU (0.93% higher than the current optimal model) in mineral segmentation, 89.49% mAP (outperforming similar models) in foreign object detection, with only 18.39M parameters and 43.06 FPS inference speed (higher than the fastest existing model). PVDNet outperforms mainstream multi-task models in accuracy and efficiency, showing good industrial deployment potential.