The invention discloses an underwater target identification method based on a convolutional neural network, comprising the following steps: firstly, simulating radiation noise of an underwater sound target; secondly, acquiring underwater target tracking beams; thirdly, acquiring a time-frequency graph of each target tracking beam, wherein all the time-frequency graphs are segmented according to fixed duration and divided into training samples and test samples; fourthly, performing data enhancement, size magnification and tailoring on the samples; fifthly, inputting the training samples provided with a label into a built convolutional neural network, performing supervised learning, and obtaining each layer parameter of the convolutional neural network; sixthly, initializing the network by utilizing each layer parameter, and obtaining the convolutional neural network with an underwater target identification function; and seventhly, acquiring radiation noise of a to-be-tested navigation target by a towed array, converting into a time-frequency graph and segmenting, inputting the segmented subgraphs into the convolutional neural network as to-be-tested samples, obtaining an identification result of each subgraph, and taking an identified target with the highest target quantity during identification as a final identification result. The method disclosed by the invention can enable underwater identification to maintain relatively high accuracy and speed under high ocean background noise condition.