An AI based framework using Aquila-optimized capsule network for radar object detection

Radar-based object detection has played a key role in current applications, such as defense, autonomous vehicles, and surveillance, with regard to object recognition and categorization of objects in diverse and multifaceted settings. However, the current radar-based detection approaches are characterized by a few issues, such as extreme inequalities between classes, suboptimal classification rates, and ineffective manual hyperparameter optimization. To address such shortcomings, we propose a deep learning-based optimized framework that improves object detection through radar signals. To overcome the issue of class imbalance, the localized random affine shadow sampling method is used to create synthetic samples of the minority classes, which is equivalent to balancing out the dataset and makes the model robust. In particular, we use a capsule network (CapsNet) with a spatial hierarchy structure and orientation, which supports a robust radar target classification. We also combine Aquila optimization (AO) algorithm with CapsNet to further enhance the performance and automate hyperparameter tuning thus introducing the proposed Aquila-optimized capsule network (AO-CapsNet) model. The experimental findings with several input multiple-output radar signal dataset prove that AO-CapsNet is much better than the current models. The proposed model achieves an improvement of 4.55% in accuracy, 2.06% in receiver operating characteristic–area under the curve, 2.11% in precision–recall area under the curve, 4.76% in Matthews correlation coefficient and Cohen’s kappa, and a 22.22% reduction in log loss compared to the baseline models. We also evaluate CapsNet with two additional optimization techniques, namely particle swarm optimization and bayesian optimization. Additionally, proximity weighted synthetic and random over-sampling examples are applied to obtain results of the proposed AO-CapsNet. To ensure reliable and generalizable performance assessment, 10-fold cross validation and paired t-test are applied to AO-CapsNet. Furthermore, to interpret the decision-making process of the proposed model and identify the most influential features, we employ explainable artificial intelligence techniques, Shapley additive explanations and local interpretable model-agnostic explanations. The integration of AO-CapsNet offers high accuracy, making it suitable for real-world radar detection applications.