Two-stage intelligent detection system of pixel-level classification and damage type recognition for membrane materials

Abstract

Damaged membrane materials would destroy the structural stress balance and have an extremely harmful impact on the appearance and safety of the cable membrane structure. The main difficulty of the membrane damage detection using the deep learning algorithm is that there are many useless redundant data due to the monotonicity of membrane images and the diversity of damage types. Therefore, a deep learning dataset of normal and damaged membrane images is established through test platform of plane bidirectional tensile membrane structure, and a two-stage intelligent detection system of pixel-level classification and damage type recognition for membrane materials is proposed. The proposed system includes two parts: (1) The normal and damaged membrane images classification method is developed based on pixel difference, which is used to determine the normal membrane material threshold. (2) YOLOv5s is trained to achieve the rapid and accurate identification of the damaged membrane dataset. The results show that the pixel percentage of normal membrane images is [0,0.0073]. Moreover, the material object detection results show that YOLOv5s offers 98.50% excellent recognition accuracy of the damaged membrane images, and verify the applicability of deep learning to the research of damage to membrane materials.