CRL-YOLO: A Comprehensive Recalibration and Lightweight Detection Model for AAV Power Line Inspections

In complex backgrounds and under severe interference, due to the relatively small size of insulator defects and dampers, existing object detection algorithms face significant challenges of high miss rates and low detection accuracy in drone-based power line inspections. To address these challenges and achieve accurate, fast detection of insulator defects and dampers, this paper proposes a novel anchor-free object detection algorithm named Comprehensive Recalibration and Lightweight - You Only Look Once (CRL-YOLO). This method uses YOLOv8n as the base framework and redesigns three key components: feature extraction, feature fusion, and detection head. First, to address the issue of inadequate feature extraction when processing multi-scale and very small objects, a C2f Star Effective Squeeze-Excitation (CSE) module was designed to enhance the network's sensitivity to multi-scale and small object features, thereby improving feature extraction capabilities. Second, to solve the problem of information loss and feature redundancy during the fusion of different level features in the feature fusion network, a Recalibration Interaction Feature Pyramid Network (RIF-PNet) is developed. This module enhances the interaction between different level features, enabling more efficient feature fusion. Lastly, considering that shared convolutional kernels can significantly reduce the number of parameters, memory usage, and computational costs, a Lightweight Unified Feature Detection Head (LUFDH) was proposed to improve the algorithm's applicability to edge computing devices. The experimental results indicate that, compared to the YOLOv8n baseline model, CRL-YOLO improves accuracy, recall, and mAP by 1%, 5%, and 4.3%, respectively, demonstrating its great potential for real-time drone inspections of power transmission lines.