Abstract: The fault caused by abnormal targets has become one of the main reasons for the outage of transmission line, resulting in great economic losses. However, due to the limited computing, communication and power resources of transmission line edge equipment, the image still needs to be sent to the data center continually (at a rate of 2 times/h) for processing, resulting in heavy load on the cloud side, high omission factor and shortage of emergency processing capacity. As a result, an abnormal target detection architecture of transmission line based on edge intelligence was developed in this paper, and an abnormal target detection model for complex background and complex shape of transmission line scene was proposed. The feature information of abnormal targets is firstly extracted with the modified MobleNetv3 network. The multi-scale target detection network YOLOv3 is then implemented to fuse high and low-dimensional feature information, benefiting the detection of abnormal targets in complex backgrounds and/or with complex shapes. Finally, an ultra-lightweight monitoring model for abnormal targets of transmission lines can be obtained based on the importance channel pruning strategy. In order to verify the performance of the proposed model, comparative experiments were carried out with a series of existing models basing on an edge device for power IoT. The experimental results have demonstrated the recognition accuracy and inference speed of the model. The proposed model has been validated to be robust and flexible, which meets the need of edge detection for transmission line abnormal targets.