CRTSII Track Slab Crack Detection Based on Improved YOLOv3 Algorithm

As one of the achievements of our country's highspeed railway, CRTSII type ballastless track, has been widely digested, absorbed and re-innovated by researchers. In the maintenance phase, however, the traditional artificial track slab crack detection method still exists some problem, such as time-consuming, low detection accuracy and difficult to detect small cracks, etc. To solve those problems, an improved YOLOv3 (You Only Look Once) algorithm is proposed. In the residual module of feature extraction network, we introduce a deep separable convolution with an inverted residual structure and SENet (Squeeze and Excitation), which reduces network parameters while appropriately deepening the depth of the network. In order to improve the accuracy of small target cracks identification, we firstly adopt the Mish activation function with better stability and accuracy, then introduce the path fusion method in the feature. Experimental results show that the accuracy of the improved YOLOv3 is 5.3% higher and the speed is 26% increase than the original YOLOv3 network. Compared with the traditional crack image processing technology, the method in this paper has better detection effect and robustness, which has a good application prospect in the detection of track slab cracks.