Insulators are an important component of power transmission lines. Using computer vision to identify the condition of insulators can improve maintenance efficiency, ensuring the stability of the line, and reducing the likelihood of fires. However, some insulators on power transmission lines are located in mountainous areas, where the complex terrain leads to variable shooting angles and significant differences in target sizes in images, making detection difficult. Additionally, visual algorithms need to be deployed on edge devices for mountainous use, requiring lighter-weight algorithms. To solve this issue, we obtained and annotated 1600 mountain insulator images as the research dataset. Then we propose the mountain insulator defect detection model LightBi-YOLO, which is an improvement based on YOLOv5. We replaced the original upsampling operation with the CARAFE lightweight upsampling operator and applied the improved BiFPN network to replace the original FPN network. Experimental results show that the LightBi-YOLO model achieves an average detection accuracy of 98.9%, representing 14% and 1.2% improvements over the Faster R-CNN model and the original YOLOv5 model. At the same time, The FLOPs of the model is 16.1, which exceeds the computational speed of the original model, making it more suitable for deployment in edge devices.