EMF-YOLO:an improved YOLO detection model for transmission line external damage hazards with enhanced multi-scale feature extraction and fusion

Abstract Transmission lines play a crucial role in power transmission, where external mechanical damage constitutes a major hidden risk for line faults. The detection of external damage on power transmission lines faces several challenges due to the complex and variable nature of their surroundings, as well as the significant size variation of targets caused by changing distances. These challenges primarily include interference from complex environmental backgrounds and insufficient accuracy in detecting small targets. This paper aims to address the issue of degraded detection accuracy resulting from background noise in complex scenarios and the loss of small-target information during the feature fusion process. Furthermore, it seeks to improve detection accuracy without increasing the model’s computational cost. This paper proposes EMF-YOLO: an improved YOLOv8 detection model with enhanced multi-scale feature extraction and fusion for detecting external damage risks in transmission line scenarios. Firstly, we propose a Diverse Branch Downsampling (DBS) module to preserve richer semantic information during feature extraction, which enhances representation capability without introducing extra computational cost. Secondly, we integrate the SENetV2 (Aggregated dense layer for channelwise and global representations) attention mechanism into the last layer of the bottleneck in the C2f module. By incorporating multi-branch fully connected layers and channel attention, this mechanism optimizes the global feature representation, thereby boosting the model’s robustness to noise and its ability to capture discriminative features. These improvements to the backbone network effectively mitigate the accuracy degradation caused by complex environments. For the neck structure, we propose a Converge-Diffuse Feature Pyramid Network (CDFPN), which optimizes cross-scale connections and incorporates an innovative diffusion mechanism. It includes a Multi-scale Feature Convergence (MFC) module that effectively integrates multi-layer features while minimizing the loss of fine-grained details from small objects during fusion. This design improves the detection accuracy for small objects and enables more effective fusion of features across different scales. Experimental results on the transmission line external damage dataset demonstrate that the improved model achieves 92.1% mAP50 and 73.3% mAP50-95, representing 3.5% and 6.5% improvements over YOLOv8m, respectively, with negligible increases in computational complexity and parameters. On the PASCAL VOC2012 dataset, EMF-YOLO achieved a 73.4% mAP50 and a 53.8% mAP50-95. Compared to YOLOv8m, there are improvements of 1.2% and 1.6% respectively. The model’s detection ability is enhanced. And through comparative experiments with mainstream existing models. The proposed EMF-YOLO exhibits superior performance compared to other existing models.