A general defect detection model for power scenarios using the improved YOLOv8

Abstract

The current centralized defect detection system faces challenges such as large data volume and poor real-time performance， underscoring the pressing need for distributed detection systems， with edge computing as a representative. In response， diverse module algorithms are devised based on the single-stage， lightweight detection model YOLOv8 to boost its accuracy in power scenarios. Firstly， the Mosaic data augmentation algorithm is improved， and a conflict relationship table is introduced to mitigate the damage to original image data information caused by traditional data augmentation algorithms and enhance the diversity of image data. Subsequently， the Res2Net module is employed to replace the original Bottleneck module， reinforcing the model’s multiscale perception while retaining its lightweight design. The adoption of the CIoU-NMS algorithm over the existing NMS （non-maximum suppression） algorithm improves the recall rate and precision of the detection model in the clustering and deduplication stage. Finally， experiments on fourteen defects in power scenarios consistently demonstrate the proposed model’s superior detection accuracy compared to the original model， accompanied by an accelerated detection speed in defect detection power scenarios.

Keywords

• deep learning
• defect detection
• data augmentation
• nms
• yolov8