Multi-Scale Component Detection in High-Resolution Power Transmission Tower Images

The UAV aerial images of transmission towers are characterized by high resolution, significant target scale variations, and specific component distribution relationships among components, challenging multi-target detection precision and speed. To address these issues, this paper combines convolutional neural network (CNN) and graph convolutional network (GCN) characteristics, proposing the globally optimized feature pyramid network (GO-FPN). First, we design the multi-branch dilated convolution (MDC) to enhance information extraction and adapt to different receptive field scales. Second, we introduce the graph-based global reasoning module (GRLite) to strengthen feature interactions between regions and improve small object detection through relational reasoning. Additionally, we propose the low information loss downsampling module (LILD), employing dense connections and space-to-depth transformations to reduce feature scales and prevent information loss, decreasing hardware consumption and accelerating inference. Finally, validation on two datasets shows that our approach achieves state-of-the-art performance in detection precision and speed.