Aiming at the problems of various shapes of foreign matter, high complexity of surrounding environment and low accuracy of existing algorithm in the detection of suspended foreign objects in substations, this paper proposes an improved Faster-RCNN object detection method. Combining the idea of feature pyramid networks and deformable convolutional networks, the method strengthens the ability of Fast-RCNN backbone to read semantic information of different scales in the input images, therefore improves its ability on detecting small objects. Using the image data in the Image Recognition Model Competition for Typical defects of substation equipment, the simulation experiment is carried out and compared with original Faster-RCNN. The experimental results verify the effectiveness of the proposed method. The improved algorithm, which has higher recognition accuracy and performs well on real detection samples, can be effectively used in the substation inspection robot system.