In general, one of the fundamental challenges in applying deep learning algorithms is the limited access to training data, which affects both the performance an‎d generalizability of these algorithms. A practical domain where this challenge is particularly significant, is the classification of moving targets with moving components in radar systems. This phenomenon, known as the micro-Doppler signature, involves identifying an‎d distinguishing targets such as helicopters, low-altitude aircraft, birds, an‎d quadrotors. In recent years, the use of deep learning algorithms in this field has grown significantly; however, data scarcity remains a major obstacle to further improving their performance. In this study, after examining several publicly available an‎d well-known micro-Doppler datasets, the Dop-Net dataset was ultimately selec‎ted due to its distinctive strengths compared to other datasets. Subsequently, a comprehensive an‎d systematic approach consisting of nine scenarios was designed to compensate for the classifier accuracy loss caused by reduced samples of this dataset. Some of these scenarios can increase accuracy under certain conditions, while under other circumstances, contrary to expectations, they may reduce classifier performance. Moreover, by eva‎luating the strengths an‎d weaknesses of each scenario, the subsequent scenarios were further refined. Finally, by enhancing an Generative adversarial network called FastGAN to generate high-quality data an‎d combining it with the Yolov11 image segmentation network for micro-Doppler feature extraction, we were able to increase the classifier’s accuracy under limited data conditions (using only 1/20 of the total data) from approximately 0.81 to 0.96, equivalent to the accuracy achieved under full data conditions. In other words, this approach compensated for an approximate 0.15 decrease in classifier accuracy under limited data conditions in the Dop-Net radar dataset. Since the design of all scenarios is independent of the dataset type, they can also be generalized to other micro-Doppler datasets.

احسان يزديان

حسين رباني

محمدرضا احمدزاده , محمدرضا تابان