Tiny bird detection and location guided by heterogeneous binocular images in transformer substation scene

Bird activities like nesting and perching in transformer substations threaten power grid stability by causing short circuits and insulation failures. Existing bird repelling devices are inefficient due to the lack of accurate detection and positioning, leading to energy waste and safety hazards from continuous operation. To address this, this paper develops a tiny bird detection and location system guided by heterogeneous binocular images for precise, targeted repulsion. For well-lit scenarios, a two-stage contextual information enhancement network is proposed. It mines multiscale context to highlight tiny bird regions, fuses context with second-stage features via channel dimension enhancement, and uses spatial attention for accurate localization. For low-light or occluded scenes, a multiscale contextual feature enhancement network processes infrared images, adopting multibranch cross-level feature fusion and combining transformer with multisize convolution to suppress background and thermal radiation interference. Additionally, heterogeneous binocular cameras are calibrated to calculate bird spatial distance, integrating detection results with spatial information to drive a laser repelling device. Experimental results in real substation environments show the system meets engineering requirements for robustness and accuracy. The detection in visible images achieves an overall average precision of $59.8$%, while the infrared detection outperforms advanced algorithms in key metrics. The spatial localization error is controlled within $4.9$%, significantly improving bird expulsion success rate and reducing energy consumption. This work provides a reliable technical solution for safeguarding power grid operation and offers valuable references for tiny object detection in complex industrial scenarios.