An Armored Target Detection Algorithm Based on KAN-RCNN

To address the issue of reduced accuracy in armored vehicle detection tasks due to reduced parameters in deep learning models, this paper proposes an improved algorithm named KAN-RCNN based on R-CNN for armored vehicle detection. The algorithm replaces traditional convolutional layers with Kolmogorov-Arnold Network(KAN) convolutional layers, reducing model parameters by 11.5% while improving detection accuracy by 6.43%. Experiments to validate the network model’s performance were conducted on 1D simulated signals, 2D MNIST, and 3D CIFAR-10 datasets. Additionally, by collecting and annotating armored vehicle images of different types, scales, and environments, an armored vehicle detection dataset was constructed and used for training and testing. The results demonstrate that replacing the traditional convolutional layers with KAN convolutional layers in the KAN-RCNN algorithm effectively reduces model parameters and improves detection accuracy. This enables the KAN-RCNN algorithm to accurately detect armored targets in complex environments, providing reliable visual support for unmanned combat systems.