Abstract: To solve the problem of underwater target feature extraction, a new network structure is proposed on the basis of convolution neural network. The framework enhances the spatial information of the feature map by introducing the feature graph multidimensional weighting layer, which makes up the loss of spatial features when entering the whole connection layer. This framework builds an endtoend network using hierarchical structure for feature extraction and classifier training, which can utilize the reverse propagation mechanism of the deep network to complete the optimization and feature extraction of the classifier simultaneously. In the simulation experiment, the classification accuracy of the network frame class reaches 78.61%, compared with other methods, it effectively improves the target recognition accuracy. The proposed framework can effectively identify underwater targets, with good recognition accuracy, and have a modular structure, without complex preprocessing.