Monitoring the real-time state of relay protections in electric power distribution substations is of significance for avoiding equipment damage and keeping the safety and stability of electrical network. Traditional state monitoring depends on regular manual inspection, which requires a great of manpower, and the character of real-time and accuracy will be restricted. To overcome this problem, a deep neural network and computer vision based monitoring strategy is proposed. First, it denoises the images collected by pan-tilt-zoom surveillance cameras, and then makes image registration according to correlations among images. Based on Scale-Invariant Feature Transform (SIFT) descriptor, the convolutional neural network can classify image and recognize the state of relay protections. Meanwhile, a modification of standard image registration framework is proposed to make model be of higher robustness under different lighting conditions. In real-world application, the accuracy of the proposed method can be higher than 99%, effectively improving the safety of substations. This work is supported by Youth Fund of National Natural Science Foundation of China (No. 61602251).