A Transfer Learning Method for Aircrafts Recognition

An effective method for recognizing aircrafts with different resolutions is proposed. Since training aircraft samples and test aircraft samples are imaging in different resolutions, different satellites and different imaging conditions, they obey different distributions. The Feature Subspace Alignment and Balanced Distribution Adaptation (FSA-BDA) method is proposed to solve this problem. Different from other transfer learning methods, it considers both spatial alignment and probability adaptation, so that, the probability distribution of the source domain data and the target domain data is as consistent as possible in the same feature space. The method first performs FSA, which maps the source domain and the target domain data to a low-dimensional common mapping space through different mapping matrices for preserving the structural information. Secondly, the BDA method is used to properly adapt the marginal probability and the conditional probability through the weight adjustment, which can leverage the importance of the marginal and conditional distribution discrepancies. This paper aims at recognizing three types of aircrafts, which are B52, F15 and F16 aircrafts. The experimental results show that the proposed method is better than several state-of-the-art methods.