In power grid inspection by an unmanned aerial vehicle (UAV), traditional deep learning algorithms may fail to achieve the high-accuracy and operate in real time owing to the characteristics of small detection area, complex background and intensive computation. To achieve accurate and fast identification of UAV power grid inspection, the detection effects of various deep learning algorithms are analyzed in detail, and an improved target detection algorithm based on YOLO v3 is proposed. The ResNet18 is adopted as the backbone network structure and a multi-scale feature pyramid is constructed. Then a deeply integrated grid inspection model is built to detect insulators via aligning the ResNet18 with the multi-scale feature pyramid, by which the detection can be executed in real time with a high-accuracy. Specifically, the mean average precision of the YOLO v3 network is 98.10%, which is increased by 6.71% over that of Faster R-CNN. Also, YOLO v3 detects up to 47.52 frames per second, 25 times and 12 times R-CNN and Faster R-CNN, respectively. The improved YOLO v3 network has better identification accuracy and detection speed. This work is supported by the Science and Technology Project of State Grid Corporation of China (No. B30970200003).