Live firing is a basic military training item of the army. The bullet hole identification and positioning system based on computer vision in the existing target detection system is widely used in this project because it is fast, accurate, safe and it has a low personnel cost. However, the image processed by the computer vision system is usually affected by the lens processing technology. In addition, the axis of the camera is not perpendicular to the plane where the measured object is located, so the image distortion of the measured object is resulted in, and the accurate positioning error of the coordinate position of the bullet hole is generated. In order to improve the accuracy of automatic target reporting system based on computer vision, a distortion correction algorithm based on convolutional neural network, which can simulate a large number of training data sets with only one template image of the chest circle target surface is proposed in this paper. When the training is completed, the distortion parameters of the image can be obtained by inputting a distorted image, and the image distortion correction can be completed by using the parameters. Compared with the traditional correction algorithm, the results show that the algorithm has a good correction effect and is conducive to improving the accuracy of the automatic target detection system based on computer vision.