Bi-YOLOv10: Sample Imbalance-Aware Feature Fusion-Based Object Detection for Power Inspection

The examination of power systems plays a pivotal role in averting power interruptions while safeguarding grid integrity as well as fostering economic resilience. Within imagery from these inspections lie diminutive target entities alongside multiscale complexities; challenges arise in distinguishing foreground from background amidst uneven sample distributions. To fortify our model's capacity for discerning intricate targets amid mitigating background noise disruptions along with environmental fluctuations necessitates amalgamating multilevel multidimensional image features. Henceforth emerges our introduction of BiFPN (Bidirectional Feature Pyramid Network) module coupled with our proposition of YOLOvlO object detection network featuring bidirectional feature fusion. Within this framework's architecture lies an integration of BiFPN replacing Concat for streamlined multiscale feature fusion efficiency; further complemented by a novel Focal-EloU loss function aimed at harmonizing contributions from high- versus low-quality samples towards enhancing model efficacy across imbalanced datasets. Experimental evaluations conducted on power system inspection datasets demonstrate superior detection capabilities exhibited by our proposed network compared against incumbent models as well as mainstream object detection networks-thus advancing secure intelligent scrutiny over power lines.