Abstract: In order to improve the identification accuracy of transformer HV bushing in infrared images and meet the needs of mobile terminal and other low-end devices for target detection network, this paper proposes an improved lightweight infrared HV bushing identification algorithm, using Tiny YOLOv3 target detection network as the basic detection network. First, through the fusion of the Convolutional Block Attention Module (CBAM) attention mechanism, the channel attention and the spatial attention mechanism are connected in series to increase the receptive field of the target detection network, while reducing network computing tasks and improving network performance. Then, GIoU loss and Focal loss are used to replace the original bounding box loss and confidence loss, thereby improving the recognition rate of the HV bushing in the infrared image and reducing the occurrence of missed and false detections. The experimental results show that compared with the original Tiny YOLOv3 network, the improved network increases mAP to 96.28%, increases F1 to 96.25%, and the weight size is 33.9 MB, less than that of YOLOv3 training network. It is better suitable for low-end equipment and provides favorable conditions for a smart substation online monitoring.

Key words: infrared images, high voltage bushing, Tiny YOLOv3, CBAM, GIoU loss, Focal loss