PMFNet: Pyramid Parsing and Multi-Scale Feature Fusion Network for TBM Muck Image Segmentation

Abstract The Tunnel Boring Machine (TBM) is the main construction equipment for modern tunnel construction. The accurate segmentation of muck images is beneficial to improving the TBM excavation efficiency and ensuring construction safety. Existing muck image segmentation methods have some problems, including inadequate feature extraction capabilities, and poor segmentation results for images with dense muck fragments. To solve these problems, we propose a muck image segmentation method based on a convolutional neural network (CNN) encoder-decoder architecture, named Pyramid Parsing and Multiscale Feature Fusion Network (PMFNet). Our method uses a Pyramid Pooling Parsing module (PPP_Block) to enhance the network's ability to obtain global contextual information, while increasing the receptive field to extract multi-scale and spatial feature information in muck images. At the same time, we use the CARAFE_Block for upsampling to restore the spatial dimensions of the feature map. To better extract edge feature information of muck, we introduce an attention mechanism (Attention_Block) to enhance the segmentation ability of the network. Finally, we construct a multi-scale feature fusion module (MSFF_Block) to further obtain texture features and global contextual features. We conducted extensive experiments on a real muck image dataset obtained from the TBM excavation scene, and the results show that PMFNet achieves an Intersection over Union (IoU), pixel accuracy (PA), and F1-Score of 60.72%, 70.37%, and 73.94%, respectively, which are better than other methods. Our method has fewer computational parameters, faster inference speed, and reaches 52 FPS, which can meet the needs of actual TBM excavation sites.