An insulator defect detection model based on improved YOLOv8

Abstract The performance of insulators in high-voltage overhead transmission lines directly impacts the safe operation of the lines. Therefore, regular inspections of insulation equipment are especially important. As the smart power grid rapidly expands, the integration of UAV aerial imaging technology and deep learning-driven computer vision algorithms has been recognized as the most widely adopted approach for evaluating transmission lines. This research proposes a novel approach for identifying insulator defects through an enhanced YOLOv8 model. To reduce computational complexity and streamline the network architecture, the GSConv is incorporated into the backbone network. Furthermore, the slim-neck design is implemented, with the GSConv and VoV-GSCSP parts being used to efficiently minimize the parameter amount of the model while improving its learning performance. Lastly, the CA attention mechanism is incorporated, embedding positional information into channel attention, which significantly enhances the accuracy of the model in recognizing insulator defects. Experimental data reveals that when applied to the CPLID dataset, the optimized model succeeds in 99.1% of [email protected] and 81.1% of [email protected]:0.95. As opposed to the original baseline model, [email protected] improves by 0.4%, [email protected]:0.95 rises by 2.1%, GFLOPs are cut by 14.8%, and model parameters decrease by 12.67%. These improvements demonstrate that the enhanced model offers better generalization and is more efficient for insulator defect detection tasks.