To make 3D reconstruction suitable for various target shapes and improve the processing speed, a 3D reconstruction method based on machine learning is proposed, which focuses on solving the "next best viewpoint" (NBV) planning problem. Firstly, the definition and calculation of NBV are given, and the discrete NBV search space is established. Then, NBV is generated, and the space is reconstructed iteratively. In addition, in order to deal with the learning problem of NBV, a classification method based on 3D convolution neural network is proposed, which considers the possible position and pose of sensor as a classification problem. The experimental results show that the reconstruction accuracy of the proposed method is better than that of the voxnet network method, which can meet the constraints better. Compared with the high-precision information gain method, the proposed method also achieves better and close reconstruction coverage. Basically, it can achieve high coverage in 4 scans for different shapes, and the reconstruction speed is about 90 times faster.