Object classification on noise reduced and data augmented micro-doppler radar spectrograms

The classification of targets is one of the most challenging tasks in radar signal processing. Classifying a target can help radar operators figure out the nature of the target, such as its source and activity. However, it is very difficult to find the labeled data necessary to develop radar target classification models. Generating a radar dataset is an expensive and time-consuming process. To address these issues, we propose a noise reduction method that can be applied to micro-Doppler radar datasets. This method is carried out by averaging the spectrograms of each class in the RadEch micro-Doppler radar datasets and subtracting pixel by pixel from each sample. RadEch dataset has also been augmented with traditional and learning-based data augmentation methods. The learning-based data augmentation method was carried out by using Generative Adversarial Networks. Raw spectrograms, augmented spectrograms and noise reduced spectrograms have been classified using 5-layer CNN, VGG-16, and VGG-19. Classification results are compared with state-of-art studies. Comparison results shows that classification on noise reduced spectrograms performs better than current state-of-art methods.