A Lightweight Insulator Defect Detection Algorithm Based on Improved YOLOv8

To tackle the challenges of slow detection speed, poor performance in detecting small objects, and large model parameters in existing insulator defect detection methods, a lightweight algorithm based on YOLOv8n is proposed. Firstly, a C2PM module is designed to replace the C2f module, reducing redundant computations and optimizing feature extraction. Then, Adown is introduced to replace the standard convolution for improving the downsampling process, enhancing the ability to retain multi-scale features. By integrating the LSCD shared convolutional detection head, efficient processing of regression and classification tasks is achieved. The enhanced model significantly reduces both the number of parameters and the computational load, making it well-suited for real-time insulator defect detection. Additionally, it performs excellently in detecting defects in small insulator targets, even in complex scenarios. Experimental results indicate that, in comparison to YOLOv8n, the improved algorithm increases mAP@0.5 by 1%, with recall and precision improving by 3% and 0.3%, respectively. Moreover, the model's parameter size is only 1MB, which is two-thirds smaller than that of the YOLOv8n network. This demonstrates that the algorithm achieves a lightweight model while also delivering a slight improvement in accuracy, realizing a dual optimization in both precision and model efficiency.