Enhanced Self-Supervised Transmission Inspection with Improved Region Prior and Scale Variation

As an important means to ensure the safety of power transmission, the inspection of overhead transmission lines requires high accuracy for detecting small objects on the transmission lines and relies heavily on the construction of large-scale datasets by using deep learning instead of manual inspection. However, transmission inspection data often involve some sensitive information and need to be labeled by professionals, so it is difficult to construct a large transmission inspection dataset. In order to solve the problem of how to effectively train only on a small amount of transmission line data and achieve high object detection accuracy considering the large-scale variation in transmission objects, we propose an enhanced self-supervised pre-training model for DETR-like models, which are innovative object detectors eliminating hand-crafted non-maximum suppression and manual anchor design compared to previous CNN-based detectors. This paper mainly covers the following two points: (i) We compare UP-DETR and DETReg, noting that UP-DETR's random cropping method performs poorly on small datasets and affects DETR's localization ability. To address this, we adopt DETReg's approach, replacing Selective Search with Edge Boxes for better results. (ii) To tackle large-scale variations in transmission inspection datasets, we propose a multi-scale feature reconstruction task, aligning feature embeddings with multi-scale encoder embeddings, and enhancing multi-scale object detection. Our method surpasses UP-DETR DETReg with DETR variants when fine-tuning PASCAL VOC and PTL-AI Furnas for object detection.