Abstract: The running state of coal mine conveyor belts is of crucial importance for safe production, and conveyor belt deviation has a significant impact on the transportation efficiency and safety of conveyor belts. Conveyor belt edge detection is an important basis for determining whether conveyor belts deviate or not. To address the problems of information loss and inaccurate extraction of conveyor belt edge lines in the traditional convolutional neural network (CNN) methods for conveyor belt edge line detection due to the difficulty in constructing long-distance dependency relationships between pixels, a conveyor belt deviation recognition algorithm based on the DETR (Detection Transformer) encoder-decoder network structure with anchor point positioning and cross-layer correction is proposed. Firstly, the anchor points of the conveyor belt edge lines are defined, and the position information of these anchor points is embedded into the self-attention module and the cross-attention module respectively. This is done to capture more comprehensive internal dependency relationships and obtain more precise image feature information, thereby enhancing the decoder’s perception ability of the conveyor belt edge lines. Secondly, a cross-layer correction strategy is added during the training stage of the model. Different supervision weights are assigned to different training stages so as to increase the influence of the later training on the whole stage and enhance the visibility of the later stage for the correction of the previous stage. In this way, the potential impact caused by cascading errors in the intermediate stage can be reduced, and the detection accuracy of the conveyor belt edge lines can be improved. Thirdly, the prior region of interest (ROI) determination is used to judge whether the conveyor belt has deviated. Finally, in order to effectively train and evaluate the model performance, 3 268 pictures of various scenes of coal mine underground conveyor belts were collected for training and testing. Through comparison and verification with the UNet, DeepLab models and the original DETR network model, the accuracy of detecting the conveyor belt edge lines by the proposed conveyor belt deviation recognition algorithm based on the DETR encoder-decoder network structure with anchor point positioning and cross-layer correction was increased by 11.15%, 9.33% and 4.73% respectively. By comparing with the on-site actually measured conveyor belt misalignment data, the conveyor belt deviation recognition accuracy correction reached 97%, and the detection speed reached 38 FPS, enabling real-time detection.