Abstract: In response to the problem of missed and false detection of insulator defects caused by small targets, multiple types, and large-scale differences during UAV inspection, a YOLOX-s object detection algorithm based on attention mechanism and multi-scales context information is proposed in this paper. Firstly, a coordinate attention mechanism is added to the backbone network to enable the network to more accurately locate insulators and their defects. Secondly, to address the issue of small target features with defects that are prone to loss, multi-scale depth-wise separable convolution is introduced into the SPP network at the tail end of the backbone network to build a multi-scale context sensitive module and to make full use of the context information. Finally, Shuffle units are used to replace the CBS stacking blocks in the feature fusion network, achieving model compression. The experiment shows that the FPS of the improved model is 26.4 frames, and the mAP value reaches 93.6%, which is 4.7% higher than YOLOX-s. Under the premise of not increasing the number of model parameters and computational complexity, better results can be achieved in multi-type and multi-scale insulator defect detection. This method has practical significance to improvement in the operation and maintenance efficiency of power inspection service.