Scholarlink

<p>Dense small object detection in complex scenes has always been a major challenge in the field of computer vision. A novel fusion detection framework is proposed to address the shortcomings of existing methods in terms of detection accuracy, computational efficiency, and feature extraction capability. This framework achieves accurate recognition of dense small objects by integrating the efficient detection capability of YOLOv7, the dynamic feature weighting characteristics of the self attention mechanism, and the multi-scale detection advantages of Single Shot Multibox Detection algorithm. Experimental verification shows that the improved model achieves an accuracy of 0.94–0.96 in vehicle detection tasks, significantly better than the traditional method’s 0.86–0.92. In complex video scenes, the F1 score remains consistently above 0.92. In addition, the average accuracy of the model on the BDD100K and MSCOCO datasets reached 95.96% and 95.23%, respectively, demonstrating excellent generalization ability. The above research results provide reliable technical support for practical applications such as drone aerial photography and intelligent transportation.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Improved YOLOv7 algorithm combined with computer vision for detecting dense small objects

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年8月25日

查看全文
  • Yun Du

摘要

Dense small object detection in complex scenes has always been a major challenge in the field of computer vision. A novel fusion detection framework is proposed to address the shortcomings of existing methods in terms of detection accuracy, computational efficiency, and feature extraction capability. This framework achieves accurate recognition of dense small objects by integrating the efficient detection capability of YOLOv7, the dynamic feature weighting characteristics of the self attention mechanism, and the multi-scale detection advantages of Single Shot Multibox Detection algorithm. Experimental verification shows that the improved model achieves an accuracy of 0.94–0.96 in vehicle detection tasks, significantly better than the traditional method’s 0.86–0.92. In complex video scenes, the F1 score remains consistently above 0.92. In addition, the average accuracy of the model on the BDD100K and MSCOCO datasets reached 95.96% and 95.23%, respectively, demonstrating excellent generalization ability. The above research results provide reliable technical support for practical applications such as drone aerial photography and intelligent transportation.