SC-HGNN: A Heterogeneous Graph Neural Network Incorporating Scattering Center for SAR ATR

Nowadays, the existing image classification networks, such as convolutional neural networks and vision transformers bear a burden of large computation, suffering from deploying on resource-constrained platforms. To this end, we propose SC-HGNN, a lightweight graph neural network incorporating scattering center features for SAR image classification. The proposed method extracts scattering center information and constructs a heterogeneous graph structure by integrating scattering centers and ordinary scattering points, effectively mitigating the issue of insufficient information in homogeneous SAR graph. In addition, we design the heterogeneous information processing layer to efficiently extract features from heterogeneous SAR graph structures, enhancing both feature representation and recognition performance. Experiments conducted on the MSTAR dataset demonstrate that, compared to existing classification networks, SC-HGNN significantly reduces computational costs, with the number of parameters (Para) averaging 1/133, FLOPs averaging 1/32, and memory cost averaging 1/32, while maintaining a recognition accuracy degradation of only 0.66%. Furthermore, compared to existing lightweight networks, SC-HGNN ensures a 0.44% improvement in accuracy with only a slight increase in training time and FLOPs, while requiring only 1/10 of the parameters and 1/3 of the memory cost, demonstrating its exceptional lightweight efficiency. Besides, the proposed SC-HGNN outperforms existing graph classification networks, which can improve recognition accuracy by 7.67%, demonstrating its superior performance in SAR target recognition.