A Transfer Learning Method for Ship Recognition in Multi-Optical Remote Sensing Satellites

In this paper, a transfer learning method of ship target recognition is proposed. This method aims at identifying unlabeled ships in high-resolution, based on the theory of transfer learning, assisting with a number of ship samples with different resolutions from different satellites. In the traditional machine learning method, training data and test data are assumed to have the same distribution. However, because of the different distributions and spaces in most cases, such as images from different satellites, the accuracy rate will decline. In this paper, we proposed a method that aligns the distributions as well as the subspace bases to solve this problem. This paper first proposed Adaptation Local Linear Embedding (ALLE) algorithm to achieve space alignment and then aligned both marginal distribution and conditional distribution by Joint Distribution Adaptation (JDA). This paper focuses on the identification of three types of ships which are destroyers, cruisers, and aircraft carriers basing on the method proposed (ALLE-JDA), using the knowledge of source samples low-resolution labeled ships to help identify the unlabeled high-resolution samples. The experimental results show that this method is better than several state-of-the-art methods.