Multiple adaptive fusion network with Mittag Leffler IoU loss for aircraft detection in remote sensing images

Aircraft detection in remote sensing is essential for airport traffic control and other applications. However, due to the influence of small objects, strong clutter, and partial occlusion, the accuracy of the aircraft detection model is severely compromised. This paper proposes a method called Multiple Adaptive Fusion Network with Mittag Leffler IoU Loss (MAML) for aircraft detection. Specifically, we first design a Mittag Leffler IoU Loss (LOSSMLIoU\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$LOSS _{MLIoU}$$\end{document}) to resolve the effects of low sample quality, distance, and aspect ratio on the model through the penalty term QMLIoU\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Q_{MLIoU}$$\end{document}. Secondly, we propose Receive Field Adaptive Spatial Pyramid Pooling (RFA-SPP), which further improves the receive field of the network by adaptively selecting the size of the pooled kernels. Finally, we propose Adaptive Feature Efficient Aggregation (AFEA), which solves the problem of information redundancy and increased computational complexity caused by direct feature fusion by automatically learning the relationship between multi-scale features. Our model achieves 96.41% mAP on the RSOD dataset and 42.64 FPS on the RTX4000 GPU, outperforming the advanced model in speed and accuracy. In addition, the results of robustness experiments on the NWPU VHR-10 and TGRS-HRRSD datasets also show the excellent portability of our method, which provides strong technical support for airport traffic control and aviation safety monitoring.