基于深度学习技术的货检图像智能识别与测试研究

A Study on Intelligent Image Recognition and Testing for Cargo Inspection based on In-Depth Learning Technology

为了提升铁路货检作业效率和质量,针对目前铁路货检工作主要依赖人工对高清图片进行查看,存在强度大、容易漏判等问题,在统计分析既有货检问题的种类、与车型的关系及所占比例基础上,采用深度学习技术,基于Faster-RCNN网络架构建立货检病害智能识别模型。经样本图片测试集测试表明,所提出的货检智能识别模型对于车顶异物、车门开启、防尘盖开启等7类问题图片的检出率接近80%,准确性超过90%,每张图片的检测速度约0.9 s,可满足现场实时、智能检测需求,从而为铁路货检向智能化发展提供有力的工具。

To improve operation efficiency and provide qualified service in cargo inspection, this paper addresses problems(such as heavy labor intensity and high rate of missing report) encountered in current cargo inspection work(which is mainly conducted by visual examination on High Definition pictures),and provides statistical analysis on categories of problems existing in cargo inspection, their relationship with vehicle types and the proportions they make up. Based on it, the in-depth learning technology is adopted and an intelligent recognition model for cargo defects inspection is constructed with the FasterRCNN network structure. Sample of tested pictures shows that the inspection rate of the above-mentioned intelligent recognition model is close to 80%, with the accuracy rate over 90%, focusing on 7 types of problems such as foreign body on the roof, irregular door opening and improper dust cover opening, and etc. According to the statistics, it takes only 0.9 second to inspect one picture. This speed could satisfy requirements on real-time inspection and intelligent testing on the spot, which provides solid support for railway cargo inspection towards intelligent development.