PID-Former: A Triple Stream Network for Real-Time Coal Flow Segmentation

Coal flow segmentation is the basis for detecting coal pile and coal flow deviation on conveyor belts, and the real-time performance and accuracy of segmentation methods are key factors affecting the above functions. However, segmenting coal flow from conveyor belts with similar color characteristics and complex lighting is a challenging task. To address the above issues, we proposed a triple stream semantic segmentation model PID-Former, which consists of P, I and D branches, and achieves real-time coal flow segmentation. Firstly, the I-Branch is designed for extracting context information based on lightweight transformer framework, and the P-Branch and D-Branch is constructed by the use of Inverted Residual Module (IRM) for detail and boundary information extracting respectively. Secondly, Context Information Fill Module (CIFM) and Edge Fusion Module (EFM) are proposed to combine different levels’ context information of I-Branch with P-Branch and D-Branch to supplement context information. At the same time, D-Branch strengthens the boundary information in features through the proposed Spatial Enhancement Block (SEB). Finally, the PID Fusion Block is designed to fuse the detail, context and boundary information extracted from P, I, D branches. Experimental results show that, PID-Former achieved the accuracy of 96.99% mIoU and 98.68% mPA on coal flow segmentation dataset, and reached a speed of 204.1FPS and 67.1FPS on the RTX3090 GPU and the embedded platform of NVIDIA Jetson TX2 respectively. Compared with the State-of-Arts, PID-Former achieved a better trade-off between accuracy and inference speed.