Light-ResKAN: A Parameter-Sharing Lightweight KAN With Gram Polynomials for Efficient SAR Image Recognition
Year: 2026
Type: article
Abstract: Synthetic Aperture Radar (SAR) image recognition is vital for disaster monitoring, military reconnaissance, and ocean observation. However, the large size of SAR images, a result of their unique imaging perspective, hinders the deployment of deep learning models on resource-constrained edge devices. Existing lightweight models often struggle to balance high-precision feature extraction with low computational requirements. The emerging Kolmogorov-Arnold Network (KAN) significantly enhances fitting capabilities by replacing fixed activation functions in traditional neural networks with learnable ones and theoretically requires only 2n+1 activation functions to approximate arbitrary functions, drastically reducing both parameter count and computational overhead. Inspired by KAN, we propose Light-ResKAN to achieve a better balance between high-precision feature extraction and low computational requirements. Firstly, Light-ResKAN fundamentally modifies ResNet by replacing convolutional layers with KAN convolutions, which allows for more intricate and adaptive feature extraction in SAR images. In addition, we leverage Gram Polynomials as activation functions, which are particularly well-suited for SAR data due to their ability to efficiently capture complex non-linear relationships. To further enhance computational efficiency, we innovatively employ a parameter-sharing strategy, where each convolutional kernel shares a set of parameters within the same channel, enabling each channel to learn unique features while significantly reducing the model's parameter count and Floating Point Operations (FLOPs). The model achieved a recognition accuracy of 99.09%, 93.01%, and 97.26% on the MSTAR, FUSAR-Ship, and SAR-ACD datasets, respectively. Experiments on MSTAR resized to 1024× 1024 show that compared to the traditional VGG16 model, our model achieves an 82.90- fold reduction in FLOPs and an 163.78- fold reduction in parameter count. In conclusion, this work establishes an efficient solution for SAR image recognition in edge computing scenarios.
Institution: National University of Defense Technology
Language: English
Subfield: Aerospace Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $1,800
OpenAlex ID: W7147557150