Machine learning for automatic target recognition

Abstract

In this thesis, machine learning such as convolutional neural network and transfer learning for automatic target recognition is researched. Recently, machine learning has shown remarkable performance in a variety of natural language processing fields such as image recognition problem, and is attracting attention in various industries and academic fields. However, in the field of automatic target recognition, it is difficult to acquire enough training data because it is military-purpose data and uses sensor data such as synthetic aperture radar and infrared image having characteristics different from ordinary image. Therefore, it is difficult to get desired performance when the conventional machine learning specific to the general image is applied. In this thesis, convolutional neural network specialized for automatic target recognition of synthetic aperture radar and infrared image are researched. It is shown numerically that the neural network shows high performance for synthetic aperture radar, infrared image even with little training data and different characteristics from ordinary image. In addition, an automatic target recognition method between heterogeneous data based on transfer learning that can supplement the insufficient training data is proposed. In the most existing research, only homogeneous data are used such as simulated data of the specific sensor and infrared data with different wavelength are used to supplement the data. This is due to the difficulty of feature learning between heterogeneous data with general machine learning methods because heterogeneous sensor data have distinctly different characteristics. In this paper, domain adversarial neural network structure, which is a recently proposed transfer learning structure, is used to enable learning in between heterogeneous data. The method is compared with the supervised learning method numerically, which is a general machine learning method for training convolutional neural network. The proposed transfer learning method shows numerical results that enables knowledge transfer in between heterogeneous data unlike supervised learning. These results suggest alternative solution to supplement little automatic target recognition data with heterogeneous sensor data.