In the visual inspection of personnel, the human position presents highly dynamic characteristics. For the situation that the robot can only capture the local information of the human body at close range leading to the inefficiency of passive vision detection, we propose an active vision method based on deep reinforcement learning. This method uses a deep convolutional network to extract image features and uses reinforcement learning strategies to train an action decision network to control the pan-tilt camera. Experimental results show that the algorithm enables the robot to turn the pan-tilt camera to achieve active face search based on the local information of the human body appearing in the image, which breaks through the limitations of the traditional method of passive detection and improves the adaptability of the visual detection algorithm in robot applications.