DyHeadNet: Insulator Defect Detection with a DYnamic Detection Head Combining Lightweight and Attention Mechanism

Considering the existing challenges in insulator defect detection, including the challenge of identifying small target defects, the limited capacity to handle complex background interference, and the low level of detection precision, this paper proposes a DyHeadNet model based on YOLOv8 for insulator defect detection. Initially, GSConv is implemented to substitute traditional convolution within the network, aiming to decrease the volume of model parameters. Following that, an efficient module named C2f-Faster is crafted to additional reduce the model's footprint. Subsequently, a novel dynamic monitoring head, DyHead, leveraging the attention mechanism, is introduced to enhance the detection capabilities for small defects. Ultimately, to precisely regulate the geometric characteristics of bounding boxes, the original model's CIOU loss function has been substituted with the MPDIoU loss function. Extensive experiments have been conducted utilizing the Insulator Defect Dataset (IDID). The outcomes demonstrate that DyHeadNet achieves an [email protected] of 96.0%. In contrast to YOLOv8n, there is a 4.8% enhancement in detection precision, while the model's parameter count and computational load are decreased by 0.5 M and 0.8 G, respectively. When pitted against state-of-the-art methods like DSA-Net and Insu-Yolo,DyHeadNet excels at identifying small targets and coping with intricate backgrounds, all while maintaining a minimal computational expense.