Damage detection of train wheelset tread based on improved SSD

Abstract

The train wheelset is a crucial part of railway vehicles, and its damage may lead to serious safety accidents. Therefore, it is imperative to detect tread damage timely and accurately. With the rapid development of deep learning, the image detection method based on a convolutional neural network (CNN) has played an important role. Single Shot MultiBox Detector (SSD) is one of the fastest algorithms in the target detection field. The algorithm has achieved excellent results in target detection, but there is a low recognition rate for small targets. Therefore, we propose an improved SSD target detection algorithm. The Original SSD algorithm is ineffective in detecting small targets with pits and cracks, so conv3-3 is selected to join the detection. We optimize convolution kernel parameters; the convolution layer contains more small target details. Compared with the original SSD, the Mean Average Precision (MAP) of tread defect is improved by 4.38%, and the MAP of small target detection is enhanced by 7.24%. This algorithm has a better performance in detection accuracy.