In view of the low accuracy of the current deep learning method in medical image registration, an unsupervised 3D convolutional neural network model for brain registration is proposed. The convolution network is used to regress the displacement field, and then the floating image is transformed through the spatial transformation layer. Then the network parameters are optimized according to the constructed loss function to realize the end-to-end unsupervised learning. By adding attention gate structure, low resolution auxiliary features are added to the connection between corresponding layers of the network to increase features and reduce background information. Compared with the unsupervised U-Net and VoxelMorph in MICCAI2012 multi-graph data, The results show that the method has higher registration accuracy and faster registration speed, and does not require expert annotation information, so it has good application potential in medical image registration.