For the infrared image-based fault diagnosis, the region of interest (ROI) needs to be selected.Due to the characteristics of many interference background and low contrast in infrared image, it isnecessary to remove the background and image segmentation to extract ROI. However, the common twovalue segmentation algorithm has the limitation of over-segmentation in the infrared image segmentation.Therefore, a method of infrared image ROI extraction based on region contrast and random forest isproposed in this paper. Firstly, the region contrast method is used to detect the infrared image significantlyto remove the interference background. Then, image segmentation is conducted by applying OTSUalgorithm in order to extract ROIinitially. Finally, aiming at realizing the optimal extraction of ROI, thethreshold of image segmentation based on the results of random forest classification is iterated andoptimized. Infrared images under 6 different conditions derived from the rotors test-bed are utilized forfault diagnosis, applying the ROI extracted by the proposed method to fault diagnosis, the accuracy of theclassification increased by 3.3 percentage points, which is more accurate than that of the artificial selectedarea.机故障诊断机制及预测预警模型研究（51674277）。