Scholarlink

<p>To overcome the limitations of traditional vision algorithms in aerial image processing, such as insufficient robustness and real-time performance, this paper proposes an image stitching method based on dynamic non-maximum suppression (NMS) and filtering mechanism. Firstly, histogram equalization is applied to improve the global contrast of the images. During the detection stage, dynamic NMS is adopted to suppress redundant feature points and prevent disordered clustering. In addition, an enhanced random sample consensus algorithm with an inner-loop iteration scheme, combined with a local bidirectional matching verification strategy, is introduced to refine Hamming distance matching results. Experimental results show that the proposed method achieves reliable feature matching under varying rotation and illumination conditions. Compared with the other four existing methods, the proposed approach achieves improvements in both matching accuracy and computational efficiency. When the resulting feature correspondences are used for image stitching, the generated global mosaic appears smoother and more natural, with additional gains observed in both RMSE and SSIM metrics.</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

Research on feature detection and matching for aerial images based on dynamic NMS and filtering mechanism

  • 期刊论文

  • 发表日期：

    2026年3月26日

查看全文
  • Weijian Gao,
  • Haijin Chen

摘要

To overcome the limitations of traditional vision algorithms in aerial image processing, such as insufficient robustness and real-time performance, this paper proposes an image stitching method based on dynamic non-maximum suppression (NMS) and filtering mechanism. Firstly, histogram equalization is applied to improve the global contrast of the images. During the detection stage, dynamic NMS is adopted to suppress redundant feature points and prevent disordered clustering. In addition, an enhanced random sample consensus algorithm with an inner-loop iteration scheme, combined with a local bidirectional matching verification strategy, is introduced to refine Hamming distance matching results. Experimental results show that the proposed method achieves reliable feature matching under varying rotation and illumination conditions. Compared with the other four existing methods, the proposed approach achieves improvements in both matching accuracy and computational efficiency. When the resulting feature correspondences are used for image stitching, the generated global mosaic appears smoother and more natural, with additional gains observed in both RMSE and SSIM metrics.