MEAFF-Net: multi-scale edge-enhanced and adaptive feature fusion network for robust object detection

The ability of existing models to accurately represent features is severely hindered in complex scenes, primarily due to the loss of critical edge information and the inherent challenges of recognizing targets at multiple scales. To mitigate these issues, conventional approaches often resort to increasing computational complexity, which in turn results in inefficient inference. In response, this paper introduces a deep learning framework that incorporates edge enhancement, multi-scale feature extraction, optimized upsampling techniques, and dynamic feature fusion to achieve more balanced and effective feature representation. First, we introduce the Multi-Scale Edge-Enhanced Feature Extraction module. It incorporates an edge enhancement strategy that separates low-frequency and high-frequency components to extract crucial edge information. Adaptive convolution is integrated to refine the representation of fine details, which in turn strengthens target perception across different scales by leveraging multi-scale feature extraction and fusion. Additionally, we propose the Feature-Driven Adaptive Reorganization module as an efficient upsampling mechanism, which improves the recovery of spatial details. For remote sensing target recognition, we further present the Residual Spatial-Channel Feature Adaptive-Feature Mixing Mechanism module. The proposed method adaptively adjusts the fusion ratio between low-level and high-level features, enhancing recognition accuracy for targets of different scales. Despite introducing slightly higher computational complexity (GFLOPs increase from 7. 20 to 11. 20, + 55%), MEAFF-Net maintains real-time inference efficiency (81 FPS at 640 640 on an RTX 3090) owing to its parallelized and content-adaptive architecture. Experimental results demonstrate that MEAFF-Net achieves detection accuracies of 83. 2%, 57. 4%, and 38. 4% on the SIMD, RDD2022, and VisDrone2019 datasets, respectively—surpassing the YOLOv11n baseline by + 3. 1%, + 2. 6%, and + 6. 1%. These results verify that MEAFF-Net offers a computationally balanced trade-off between accuracy and inference speed, showing strong potential for real-world multi-scale object detection applications.