Analyzing DSIAC ATR algorithm development database utilizing transfer learning

Abstract

For military applications, recognizing the targets with a good accuracy is a vital skill. In the literature there are many machine learning-based works on target recognition in visible spectrum, since there are massive RGB datasets. However, it is very important to have the same capability in infrared spectrum for military applications. Because of that reason DSIAC database, which has both visible and infrared images of the same targets is introduced first. Then a straightforward and efficient transfer learning-based ATR algorithm is proposed. Each step of the transfer learning process is explained in detail. The proposed transfer learning algorithm is tested with many challenging scenarios of DSIAC database. We extract valuable results how the ATR performance depends on range, wavelength and time changes. We also test ATR capability of our proposed model against extensive data. At the end we achieve very satisfactory accuracy scores thanks to the power of transfer learning.