In order to solve the problem of insulator missing detection and inaccurate positioning caused by small proportion of insulators and complex background in infrared power image, a novel insulator detection network: Feature Selection YOLOv3(FS-YOLOv3) was proposed. The proposed FS-YOLOv3 added pyramid feature attention network to the top-down sampling process of the original pyramid shaped YOLOv3 network. The pyramid feature attention network calculated the feature weight matrix based on the high-level semantic feature map of YOLOv3, and used the feature weight matrix to filter out the redundancy of low-level detail features of the network. Finally, the low-level feature map and the high-level semantic feature map after feature filtering were connected in series to obtain the feature map with both accurate insulator detail information and rich high-level semantic information. The experimental results show that the detection accuracy of the proposed method is better than that of the original YOLOv3 network, and retains the good real-time characteristics of the original network.