The invention relates to a three-dimensional image matching method based on a convolutional neural network, which comprises the following steps: acquiring a plurality of three-dimensional matching image pairs and corresponding real disparity, and taking the three-dimensional matching image pairs and the corresponding real disparity as a data set; Constructing a convolutional neural network, and selecting a linear correction unit (RELU) function for activation; Training the convolutional neural network by adopting a back propagation algorithm, and determining a network error function and a learning rate; Through the calculation of the convolutional neural network, the network outputs a matching cost space diagram of the left and right image blocks; And performing matching cost aggregation,disparity selection and disparity refinement on the cost space graph, and selecting a pixel point with the minimum cost as a matching point to obtain a final disparity map. The image can be directly used as network input, the overall disparity map obtained through the convolutional neural network matching algorithm is relatively smooth, a relatively good matching effect can be achieved in a non-texture region and a depth value abrupt change region, and relatively good robustness is still achieved for an image pair with illumination change and incomplete correction.