Multi-order information aggregation network for remote sensing small target detection

In critical fields such as intelligent transportation and national defense security, remote sensing small target detection technology plays an essential role. To effectively overcome the complexity of remote sensing scenes and the weak response of small-scale targets, this paper proposes a lightweight Multi-order Information Aggregation Network (MIANet). MIANet mainly consists of two parts: Cross-spatial Multi-order Information Aggregation Module (CMIAM) and Multi-dimensional Information Enhancement Module (MIEM). Inspired by the research on multi-order interactions in game theory within deep learning, CMIAM can aggregate low-order, mid-order, and high-order information, effectively improving the detection accuracy of small targets in complex remote sensing scenes. Based on the design philosophy of manifolds of interest, MIEM can effectively remove redundant information, and MIEM utilizes a three-branch structure to capture cross-dimensional information interaction, enriching feature representation and achieving the effect of information enhancement. We have validated the performance of our model on multiple remote sensing small target datasets including VEDAI, DIOR, NWPU-VHR10, MVRSD, and SIMD, and achieved excellent results. In particular, for the lightweight MIANet, the accuracy metric $\text{m}A{P}_{50}$ reached 73.7% on the VEDAI dataset, surpassing the current SOTA method SuperYOLO for remote sensing small target detection on a single modality. On the NWPU-VHR10 dataset, MIANet outperformed SuperYOLO by 2.1% in the $\text{m}A{P}_{50}$ metric and FFCA-YOLO by 2.2%. On the DIOR dataset, with a parameter count of 9.79M, MIANet achieved an $\text{m}A{P}_{50}$ metric of 81.3% and an $\text{m}A{P}_{50:95}$ of 60.9%, which demonstrate that our model exhibits strong robustness characteristics. Our code will be made publicly available on https://github.com/Liro-o/MIANet.