Abstract: Aiming at the requirements of high precision and high efficiency detection of rolling contact fa- tigue testing machine, an improved Otsu algorithm based on automatic weighted threshold was proposed. The algorithm took the probability of occurrence of the background as the weight of the Otsu segmentation threshold, and made the segmentation threshold close to the lower left edge of the histogram. Firstly, the image of roller contact surface was obtained by region location of roller image. Secondly, the detection ac- curacy rate, recall rate, harmonic mean and running time of different filters were compared. Fast guided fil- ter with short running time and large harmonic mean was selected to filter the image. Finally, the misclassi- fication error of the image was calculated to verify the effectiveness of the detection algorithm. Compared with Otsu, weighted object variance, valley-emphasis and neighborhood valley-emphasis methods, the ex- perimental results show that the misclassification error of the proposed algorithm for the test image is close to 0, and it has better effect in rolling contact fatigue defect detection.