Para-YOLO: An Efficient High-Parameter Low-Computation Algorithm Based on YOLO11n for Remote Sensing Object Detection

Remote sensing object detection has shown considerable application potential in environmental monitoring, disaster management, and urban planning. However, technical challenges persist due to background complexity, varying object scales, and interclass similarities. Moreover, the increasing demand for real-time onboard image processing imposes stringent requirements on both algorithmic complexity and detection accuracy. To overcome these challenges, we propose Para-YOLO, an efficient and cost-effective algorithm consisting of a novel feature fusion network and three innovative modules. RLFFN specifically addresses the prevalence of small and medium objects in remote sensing imagery. By utilizing the intermediate layer of the feature fusion network as the aggregation-diffusion layer, it mitigates the feature degradation caused by consecutive upsampling in Feature Pyramid Networks. The MPMS module not only reduces computation, but also increases the number of parameters and improves the model’s ability to extract local multiscale features. The LPSM module enhances the perception of fine-grained spatial features. The Center Information-Guided Aggregation and Diffusion Module serves as a contextual interaction hub within the RLFFN. Compared to YOLO11n, Para-YOLO has 2.5 times more parameters (6.5 M versus 2.6 M), while the computation increases by only 1 GFLOPs (7.3 G versus 6.3 G). Initially, Para-YOLO is evaluated on the SIMD dataset, achieving 83.3% in terms of mAP@50. This performance surpasses YOLO11n by 3.4% and outperforms classical and state-of-the-art methods. The generalization capability of Para-YOLO is further validated on three additional datasets, demonstrating its substantial potential for deployment on edge platforms such as artificial satellites and uncrewed aerial vehicles (UAVs).