钢轨表面伤损的细粒度图像识别

Fine Grained Image Recognition for Rail Surface Defects

摘要:

基于智慧工务对钢轨轨面状态和伤损的精准识别、定量化修理等需求，结合深度学习和机器视觉，提出了钢轨轨面伤损细粒度图像识别与量化方法。通过采集轨面状态和伤损图像并实现伤损细粒度标注，建立轨面伤损RD-1 094数据集，其中的目标密度达到了每图22.9个。建立轨面伤损深度学习目标检测算法，通过对RD-1094数据集的训练和学习，实现了对0.5~30mm的剥离掉块、波长20~200mm的波磨等轨面伤损及其各自发展阶段特征的识别，达到毫米级细粒度。算法对单双排波磨、细小密集和轻重伤剥离掉块、单支和成片疲劳裂纹等能较好的兼容性，可以实现轨面光带形位、伤损尺寸、轻重伤总数、分布面积、波磨波长等状态和伤损的量化评估。

关键词:轨道交通;钢轨;表面伤损;目标检测;图像识别

Abstract:

Based on intelligent engineering technology， precise identification of track defects and quantified repair issues， combined with deep learning and computer vision techniques， a method for fine-grained image recognition and intelligent evaluation of rail surface defects has been proposed. Rail surface Defects dataset （named RD-1 094 dataset） was established by collecting images of rail surface defects and realizing fine-grained annotation of defects. The target density of the dataset reaches 22.9 targets （defects） per image. Moreover， a target detection algorithm with the ability of deep-learning for rail surface defects was established. Through training and learning to the RD-1094 Dataset， it achieved millimeter-level fine-grained recognition， such as realizing the recognition of spalling with size in 0.5 ~ 30 mm， corrugation with a wavelength in 20 ~200mm and other rail defects with their own growth situation. The algorithm has good generalization compatibility on single or double row corrugations， small or dense spallings in different levels， fatigue cracks with a single piece or with obscure distribution on the rail surface. It can measure the shape and localization of contact band on the rail surface， the size of the defects， the total number of defects in different levels， cracking area， the wavelength of corrugation and other quantitative evaluation metrics.

Key words:rail transit;rail;surface detect;target inspection;image recognition