Underwater acoustic target recognition (UATR) in ship acoustic data poses significant challenges. Today, deep learning methods is widely employed to extract various types of information from underwater audio data. This paper explores the application of one-dimensional and two-dimensional convolution methods for detection. The raw acoustic data captured by hydrophones undergoes necessary pre-processing. Subsequently, regions of interest (ROI) that contain ship-emitted noise are extracted from spectrogram images. These regions are then fed into convolutional layers for model validation and classification. One-dimensional methods have faster processing time, but two-dimensional methods provide more accurate results. To significantly reduce the computational costs, in this paper, three effective algorithms based on deep learning for object detection are presented, which can be found by searching for the most informative features from the labeled data and then continuous training of the model of integration. New labeled samples with pre-labeled samples at each epoch will increase the accuracy of recognition and reduce losses. Through the combination of diverse pre-processing steps and modified deep learning methods, the proposed method achieves a recognition accuracy of 97.34% when tested on a dataset consisting of four types of ship-radiated noise. The method demonstrates superior performance compared to other deep learning methods.