Abstract: The existing belt tear detection based on computer vision has the disadvantages of poor imaging quality, weak generalization ability, and high hardware cost, and a belt tear detection based on full convolutional neural network is proposed. The hardware is composed of a line structured light and an edge computing device. Firstly, the can-didate laser stripe is extracted by the extreme value method, and the complete laser stripe is extracted after the erroneous pixels are reconstructed by the nearest neighbor method. Secondly laser stripe is segmented by one-dimensionalized full convolutional neural network. Finally, the segmentation result is back-projected to pixel coordinate system, and then the result in pixel coordinate system is mapped to world coordinate system by the line structured light, so the tear position is located and the dimensions are measured. The experimental results show that: This stripe detection method is better than the steger method and the the grayscale-gravity method. The stripe segmentation based on one-dimensionalized U-net is satisfactory, which's DICE is equal to 94.71% and MIOU is equal to 94.70%, then the recall rate of tear detection is 96.09% and the accuracy rate is 96.85%.