基于并联卷积神经网络的高速铁路车体图像差异分类算法

Change Classification Algorithm of High-speed Train Body Images Based on Parallel Convolutional Neural Networks

对高速铁路的车体进行安全检查,是高速铁路安全运行的重要保障。但是由于高速铁路车体结构复杂,采集环境多变,导致基于机器视觉的传统检测方法很难提取到车体图片的正确特征。利用深度学习方法,采用卷积神经网络,建立一个并联的差异分类模型,用于检测车体差异部位是否发生异常。为提高识别的准确率,提出一种多形状训练方法;同时,针对此异常检测任务设计合适的损失函数,并加入距离度量的限制项。此外,在并联模型的基础上,建立一种复合并联模型,进一步提升模型性能。试验表明,该模型能够很好地克服光照、污渍、标记等伪异常,正确提取到图片对的差异信息,并对差异信息是否异常做出准确的判断。

Safety inspection on the body of high-speed trains is an important guarantee for their safe operation. Conventional detection methods based on computer vision can hardly extract the correct features of the body images, especially for the images of high-speed trains which have complex body structure and variable acquisition environment. In this paper, a parallel model was applied to detect whether changed images of corresponding components were abnormal by using deep learning method based on convolutional neural networks(CNNs). To improve the classification accuracy, a multi-shape training method was introduced, and an appropriate loss function was designed for the task with adding a limitation of distance measurement. In addition, based on the parallel model, a composite parallel model was proposed to further improve the performance of the model. Extensive experiments demonstrate that the proposed methods perform well and overcome the influence of false abnormities such as illumination, stain and marks. The model can correctly extract the change information of image pairs and determine the abnormality of the change information.