Abstract: The marine environment is complex and changeable, and the target recognition accuracy of a single underwater sensor can not meet the performance requirements of the system. Multi-source sensor fusion is an effective method that can improve the target recognition rate. It has received extensive attention and research. Underwater data is noisy and has high dimensions, and direct data fusion can not get better results. For multi- scene underwater multi-source test data, denoising autoencoder and multiple dimensionality reduction methods is used for multi-angle feature-level fusion. First, the denoising autoencoder is used to remove noise and reduce the data dimension, and extract new features from the source data. Then, the data cascade method is used for multisource data fusion for new features. The fusion methods include principal component analysis, independent component analysis and isometric mapping. Comparative experiment results in different scenarios show that the proposed method gets better classification results, and principal component analysis can achieve a higher target recognition rate.

Key words: underwater multisource data, feature-level fusion, denoising autoencoder, dimensionality reduction method, data stacking method