Abstract: A deep convolution neural network method using adaptive enhancement learning is proposed to extract underwater acoustic target line spectrum. Firstly, a multi-scale ConvNeXt operator is constructed. Then a sliding window deep convolution neural network model (SwDCNN) for extracting line spectrum from Lofargram is established based on the operator. Finally, an adaptive enhancement learning criterion covering loss function, learning rate update and model iterative optimization is designed and used for SwDCNN model training. The numerical simulations and sea trial data show that the proposed method has four advantages: (1) The structure parameters of SwDCNN model are configured according to the line spectrum extraction requirements, which can enhance the multi-resolution mining ability of Lofargram features. (2) Under the condition of large-scale data, the model training effect of progressive accurate fitting can be achieved, and the model convergence effect can be improved. (3) The line spectrum under complex conditions such as low SNR, interruption, bending drift, uneven thickness, clustered nearest neighbors and dense distribution can be effectively extracted by the SwDCNN model. In terms of recall, precision, false alarm, LLA (Line Location Accuracy) and LAA (Line Amplitude Accuracy), the SwDCNN model is superior to other deep neural network methods used in this paper. (4) Compared with the traditional and other common deep neural network methods, the line spectrum minimum detectable SNR of the SwDCNN model is reduced by more than 5 dB and 2 dB respectively. The SwDCNN model has stronger line spectrum extraction ability and better comprehensive effect in actual complex scenes.