| Title |
Performance Improvement of Space-Target Classifier Based on Convolutional Neural Network Using Angle of Incidence |
| Authors |
김준선(Jun-Seon Kim) ; 서동욱(Dong-Wook Seo) |
| DOI |
https://doi.org/10.5370/KIEE.2024.73.9.1551 |
| Keywords |
Angle of Incidence; Convolutional Neural Network; Space-Target; Micro-Doppler Signature; Micro-Motion |
| Abstract |
Due to the characteristic micro-motion of space-targets, a micro-Doppler effect occurs when a radar observes the space-target. The two-dimensional micro-Doppler signature image, which well represents the micro-Doppler effect, is used as a feature in classification using convolutional neural network in many literature. The angle of incidence of electromagnetic waves incident on a space-target is one of the information that can be obtained during radar observation. In this paper, we propose a method to improve the performance of the space-target classifier by using this angle of incidence as a feature in training a convolutional neural network model. The angle of incidence is input to the fully connected layer by concatenating it with the feature maps that are the output of the convolutional layer, and this was applied to ResNet-18 and a simple convolutional neural network model. Although the performance improvement in ResNet-18 was small compared to the simple model, it was clear in all cases, and classification accuracy was especially improved at low SNR. When the weight of the angle of incidence was set large, the F1-score showed a larger increase than when the dwell time was doubled, showing that it can be efficiently applied to space-target classification where decision-making must be completed within a short time. |