The invention provides a furnace mouth image preprocessing method. The method includes acquiring the furnace mouth flame image and removing the invalid regions on the periphery of the furnace mouth flame image to obtain an initial ROI; detecting the initial ROI and judging whether the invalid region with the longitudinal penetration exists or not; if the invalid region with the longitudinal penetration exists, segmenting the initial ROI into a plurality of sub-graphs; and removing the invalid sub-graphs in the sub-graphs to accurately extract the ROI. For the problem of image noise caused by the shielding of a furnace mouth area by a fireproofing door and a smoke cover under the complex light condition of a converter working site, the various irrelevant content and interferent objects canbe removed, and an effective ROI can be obtained, the recognition accuracy of the state of the converter can be obviously improved, the method has the advantages of being easy to realize and small incalculation amount, the requirements of real-time performance processing can be met, and the effectiveness of input images in the depth learning model can be guaranteed.