As the two edge pixels of a heavy rail image is not identical, the classical edge operators are difficult to achieve the edge extraction and segmentation. Therefore, this paper proposed a new algorithm to enhance and extract images. A strong contrast stretching algorithm was used to stretch the rail surface and the background differently, enhance the edge information and weaken the background information. Then, the maximum variance method was taken to select the optimal threshold to implement the binarilization. Finally, the recursion connected domain marker algorithm was used to locate the pixel coordinates of edge to achieve the image segmentation. 30 images were chosen to a discretion experiment, and results indicate that the gray features of image edge are enhanced clearly, surface textures and false edges are restrained availably. Moreover, the pixel width fluctuating range is reduced from -0.64% to 0.34%. With the discrete pretreatment algorithm via addressing global array of a register, the segmentation time has been decreased to 10.165 s. The algorithm is better than the classical edge operators in the precision, correctness and the timeliness and is more suitable for on-line detection systems.