MambaRetinaNet: Improving remote sensing object detection by fusing Mamba and multi-scale convolution

Object detection in remote sensing imagery is a fundamental task in Earth observation applications. However, its accuracy faces significant challenges due to the influence of complex backgrounds and multi-scale object features. Despite advancements in the methods utilizing convolutional neural networks (CNNs) and self-attention, they encounter two fundamental challenges. CNNs, constrained by their local receptive fields, struggle to capture adequate global feature representations. Conversely, while self-attention excels at modeling long-range dependencies, its high computational complexity hinders practical application on high-resolution remote sensing imagery and can degrade the representation of local details. To resolve these challenges, this article proposes a novel detection model named MambaRetinaNet, which fuses multi-scale convolution with the Mamba architecture. We designed a Synergistic Perception Module (SPM) to efficiently model global information while enhancing local feature retention. Furthermore, we introduce an Asymmetric Feature Pyramid (MambaFPN), which optimizes the feature pyramid network through a differentiated processing strategy to achieve a balance between detection accuracy and computational efficiency. The experimental results indicate that MambaRetinaNet demonstrates significant advantages on four benchmark remote sensing datasets. Specifically, the mean Average Precision (mAP) on DOTA-v1.0, DOTA-v1.5, DOTA-v2.0, and DIOR-R are 80.03, 70.21, 57.17, and 71.50, respectively — representing an average improvement of approximately 11 over the baseline. While introducing a slight increase in the number of parameters, MambaRetinaNet nonetheless substantially reduces the computational cost, lowering the FLOPs by nearly three times compared to the baseline, and maintains the lowest FLOPs among competing methods. These results highlight the effectiveness and practical value of the proposed method in complex remote sensing scenarios.