An effective feature for crack detection on train wheel surface

Automatic crack detection on train wheel surface is highly crucial for safety and labor-cost efficiency. Algorithms have been explored to cope with the image based crack detection problem, but mostly for concrete structure or pavement surface. It´s a tougher problem on train wheel surface because of noise such as scratch, rust, dirt, shadow and son on which are unlikely to be denoised by simple image processing methods or image transform algorithms. In this paper, we treat this mission as an object detection problem. A LoG gradient based method is proposed to weight the pixels within the Haar-like feature mask region. We call it LoG Weighted Haar-like feature (LWH). With this, a more precise description of the crack sample images is achieved. An evaluation is performed to show the effectiveness of our method.