Perturbation-Based Analysis of the R(2+1)D Network and a Deformable Extension for Radar-Frequency Image Classification

Target identification is one of the fundamental challenges in radar systems, involving the classification of detected objects based on their type (e.g., human, vehicle) and motion state. While traditional approaches rely on analyzing variations in the target’s motion vector parameters, advancements in artificial intelligence have enabled more accurate and efficient methods. This paper presents a comparative study of deep learning architectures for the classification of radio-frequency images. We investigate two custom hybrid models: a residual (2+1)D convolutional neural network combined with a long short-term memory layer, and our proposed residual (2 deformable +1)D variant, which integrates deformable convolutions to better model nonrigid motion. These convolutional neural network based frameworks are benchmarked against transformer-based approaches, specifically the TimesFormer architecture and a vision transformer pretrained using the video masked autoencoder method. To interpret and evaluate the decision-making processes, we employ two distinct explainability techniques. For the convolutional neural network based models, the spatio-temporal extremal perturbation method is utilized to reveal key regions influencing classification. For the transformer-based models, we analyze their inherent attention maps to gain insight into spatio-temporal areas of focus. This dual approach allows for a comprehensive understanding of how different model architectures process radio-frequency imagery for target classification.