Instance-Level Detection and Region Partition of HPT Blade by Slot Localization

Accurately partitioning a high-pressure turbine (HPT) blade is very important for deciding the severity of the damage occurring on an HPT blade. Although it is important, there is no autonomous HPT blade partition method. A direct way is to segment an HPT blade by the state-of-the-art (SOTA) instance segmentation method trained with abundant labelled images. However, the segmentation-based strategy is restricted by the number of pixel-level labels, which cannot meet the different partition standards of variant aeroengines. In this paper, we study how to partition an HPT blade without pixel-level labels. Our method is based on slot localization, dubbed as SL-BP, which can decouple the HPT blade partition with instance segmentation, and make SL-BP can be easily applied on a new type of aeroengine without relabeling. Given the indexes of localization slots, SL-BP can easily generate promising HPT blade partition results. To improve the blade partition results further, we propose “smart YOLACT”, which is pretrained with pseudo labels generated by SL-BP and finetuned with a small number of pixel-level labelled images. We build an HPT blade partition dataset, named BPD, with 107 images of the HPT blade. Compared with the instance segmentation-based method that is trained with pixel-level labels, SL-BP can achieve comparable performance only with bounding box annotation on slots of HPT blades. Besides, finetuned with a half number of labelled images, our “smart YOLACT” even outperforms YOLACT trained with all the labelled images.