Real-Time Semantic Segmentation Model Combining CNN and Transformer to Detect Belt Runout

Along with the development of science and technology, the belt conveyor has been applied on a large scale. Belt runout is one of the common faults of belt conveyors. To address the problem of belt deflection detection, this paper constructs a binary belt segmentation dataset BSG containing two categories of belt and background and designs a real-time semantic segmentation-based belt deflection detection method. Transformer has excellent global modeling capability but lacks spatial location bias, and CNN has spatial location bias and has advantages in local information extraction. In this paper, a real-time semantic segmentation network LCTNet combining CNN and Transfomer is proposed. LCTNet contains the shallow feature extraction module LSBM, the deep feature extraction module LDBM, and an efficient Transfomer Block (SET). Experiments on the CamVid dataset as well as the BSG dataset show that the LCTNet balances accuracy and real-time performance in a good way.