BCM-YOLO: An improved YOLOv8-based lightweight porcelain insulator defect detection model

Abstract

Porcelain insulator is an important component of power transmission systems, and its condition detection is essential to ensure safe operation of the power grid. Nevertheless, it is difficult for existing detection models to effectively solve the contradiction between detection accuracy and resource consumption. To address this issue, a high-precision lightweight insulator defect detection model (BCM-YOLO) based on an improved YOLOv8 is proposed. Firstly, bidirectional feature pyramid network (BiFPN), with a simplified bidirectional information flow mechanism, is employed to replace the path aggregation network with feature pyramid network in YOLOv8 to alter the feature fusion mode, thereby reducing the model size. Secondly, a cross-stage partial Bottleneck with 2 convolutions partially replaced by a context-guided block (C2f_CG) structure with parameter sharing is designed using the improved context-guided block to optimise the cross-stage partial Bottleneck with 2 convolutions (C2f) modules, thus further decreasing the number of model parameters. Finally, multiscale dilated attention is introduced into the BiFPN network to enhance the perception ability of different scales of features to improve the detection performance. Experimental results indicate that compared to YOLOv8s, the BCM-YOLO model reduces the number of parameters by 50.5%, lowers floating-point operations by 31.3% and increases mean average precision at intersection over union = 0.5 (mAP0.5) by 2.8%. The proposed model not only improves detection accuracy but also decreases parameter counts, making it more suitable for deployment on edge devices.