Weakly Supervised Detection of Overhead Contact System Bolts Based on Improved Proposal Cluster Learning

Bolts are critical components of overhead contact systems (OCS), and their timely detection of looseness or absence is of great significance in maintaining the structural safety and stability of key components of OCS. In response to the characteristics that bolts in the picture are small in size, numerous in quantity, and irregular in shape, making them difficult to annotate, we propose a weakly supervised detection method for bolts based on improved proposal cluster learning, which can accurately locate bolts with only image-level annotations. Specifically, image-level annotations only require labeling the classes of the targets in a given image (e.g., using a binary vector with 0s and 1s to indicate the existence of a particular type of bolt in the image). Compared to instance-level annotations (e.g., bounding boxes) required by fully supervised models, this approach significantly reduces the difficulty of annotation and saves labor costs for creating bolt datasets. Through optimization experiments such as NMS threshold, multi-scale training, and IoU threshold, as well as the design of the weighted loss function, the proposed method achieves detection results with mAP of 27.4% and Corloc of 48.7%.