YOLOv7-b: An Enhanced Object Detection Model for Multi-Scale and Dense Target Recognition in Remote Sensing Images

To address the challenges of dense object distribution, scale variability, and complex shapes in remote sensing images, this paper proposes an improved YOLOv7-b model to enhance multi-scale target detection accuracy and robustness. First, deformable convolution (DCNv2) is introduced into the YOLOv7 backbone to replace the standard convolutions in the last two ELAN modules, thereby providing more flexible sampling capabilities and improving adaptability to irregularly shaped targets. Next, a Bi-level Routing Attention (BRA) module is integrated after the SPPCSPC module, employing both coarse- and fine-grained routing strategies to focus on densely distributed targets while suppressing irrelevant background. Finally, training and evaluation are conducted on the large-scale DIOR remote sensing dataset under unified hyperparameter settings and evaluation metrics, allowing a systematic assessment of the overall model performance. Experimental results show that, compared with the original YOLOv7, the improved YOLOv7-b achieves significant enhancements in Precision, Recall, mAP@0.5, and mAP@0.5:0.95, with mAP@0.5 and mAP@0.5:0.95 reaching 85.72% and 66.55%, respectively. Visualization further demonstrates that YOLOv7-b provides stronger recognition and localization for densely arranged, small-scale, and morphologically complex targets, effectively reducing missed and false detections. Overall, YOLOv7-b delivers higher detection accuracy and robustness in multi-scale remote sensing target detection. By combining deformable convolution with a dynamic sparse attention mechanism, the model excels in detecting highly deformable objects and dense scenes, offering a more adaptive and accurate solution for small-target detection, dense target recognition, and multi-scale detection in remote sensing imagery.

@article{Song2025,
  title     = {YOLOv7-b: An Enhanced Object Detection Model for Multi-Scale and Dense Target Recognition in Remote Sensing Images},
  journal   = {International Journal of Advanced Computer Science and Applications},
  volume    = {16},
  number    = {2},
  year      = {2025},
  publisher = {The Science and Information Organization},
  author    = {Yulong Song and Hao Yang and Lijun Huang and Song Huang},
  doi       = {10.14569/IJACSA.2025.0160225},
  url       = {https://doi.org/10.14569/IJACSA.2025.0160225}
}