深度学习在接触网定位器缺陷检测中的应用

Deep learning applied to defect detection of contact line locator

高速动车接触网运营安全的需求使得接触网关键零部件的缺陷自动检测成为一份有意义的工作。针对接触网巡检图像的定位器缺陷检测问题,本文提出了一种基于图像深度表示和直线检测的目标检测一体化算法。该算法采用选择搜索算法获得定位器在图像中可能存在的备选区域,利用深度卷积神经网络计算图像的深度特征,通过多任务学习的算法求得定位器的局部区域。随后,利用Canny边缘提取和Hough直线检测的方法在局部区域内精确检测定位器直线。针对接触网巡检图像的实际应用场景,对该算法在不同场景下进行验证,试验结果表明,该算法可以有效解决实际场景下的定位器缺陷检测问题。

Aiming at the locator defect detection of contact line inspection image,this paper proposed a detection integration algorithm based on image depth representation and line detection.This algorithm adopted selective search algorithm to obtain the possible alternative regions in image where locator may exist.The deep features of image were calculated by deeply convolutional neural network and local region of locator could be obtained by multi-task learning algorithm.Subsequently,by using Canny edge extraction and Hough line detection method,locator line could be exactly detected in local region.According to the practical application scene of contact line inspection image,this algorithm has been tested by several locator detection tasks at different scenes.The experiments proved that this algorithm could solve the detector defect detections problems effectively in complex environment.