Geometric highlight time delay estimation (GHTDE) is the major technique of underwater target size recognition. The distortion of underwater target echo signal caused by ocean ambient noise and multipath effect in underwater acoustic channels can affect the accuracy of GHTDE. In order to complete the task of GHTDE under signal distortion, a deep learning method is proposed in this paper named GHTDE network. By initializing the weights of the convolutional neural network (CNN) layer with the complex conjugation of the reference signal, the GHTDE network has the same function as matched filter (MF). The auto-encoder (AE) structure combining with CNN layer enable the GHTDE network to suppress signal distortion. Meanwhile, we propose two kinds of evaluation indicators, which could evaluate the time delay estimation performance respectively from time delay position recall rate and time delay estimation mean absolute error. The experimental results indicate that GHTDE network outperforms classical MF in terms of time delay estimation performance and robustness to signal distortion.