Underwater target recognition method based on t-SNE and stacked nonnegative constrained denoising autoencoder

Abstract

Underwater targets recognition is a difficult task due to the specific attributes of underwater target radiated noises, low signal to noise ratio and so on. In this paper, the input data optimization method and recognition model were researched. The underwater target radiated noise spectrum was chosen as the original feature. The t-distributed stochastic neighbor embedding (t-SNE) algorithm was used to reduce the dimensionality of the original spectrum segments divided by frequency. The optimal features can be obtained by analyzing the separability. Then the stacked nonnegative constrained denoising autoencoder (SNDAE) model was established to recognize the optimal features. The experimental signal spectra were processed by above methods. The results show that the recognition accuracy of SNDAE is higher than that of other contrastive methods. And the frequency of input band with the highest recognition accuracy is approximately the same as that with the best separability based on t-SNE, indicating that the above method can improve the recognition accuracy and efficiency.