Scholarlink

<p>Target detection algorithms have difficulties accurately recognizing specular image regions, mainly owing to light reflections and interference from environmental factors. Despite recent advancements in research, two major issues remain: First, prior studies have only considered the segmentation of specular and non-specular regions while ignoring the similarity of the content within the specular region. Second, these methods rely on a large number of parameters to achieve good segmentation performance. To address these issues, in this study, we propose a frequency domain fusion network (FDFNet). In addition, to better establish content similarities, we introduce the frequency–time blending module to extract the combined features of the red, green, blue, and depth frequency domains. We also propose a multiscale fusion enhancement module to improve multiscale semantic features. In addition, we combine a semantically complementary comparative learning refinement approach combining comparative learning and knowledge refinement, which divides knowledge into foreground and background. For both teachers and students, the same foreground or background features are brought closer together, while different foreground or background features are moved further apart. The FDFNet-S student model is improved under the guidance of the FDFNet-T teacher model, resulting in FDFNet-S*. Our design compresses the number of parameters and floating-point operands from 164.69 million and 78.86 billion in FDFNet-T to 17.41 million and 7.76 billion in FDFNet-S, respectively. The experimental results confirm the effectiveness of the proposed methodology and framework. The code is available at <a href="https://github.com/sfsffafasf/FDFNet">https://github.com/sfsffafasf/FDFNet</a>.</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

FDFNet-S*: frequency domain fusion networks for RGB-D mirror segmentation by contrastive knowledge refinement

  • 期刊论文

  • 发表日期：

    2025年4月10日

查看全文
  • Chao Li,
  • Caie Xu,
  • Wujie Zhou

摘要

Target detection algorithms have difficulties accurately recognizing specular image regions, mainly owing to light reflections and interference from environmental factors. Despite recent advancements in research, two major issues remain: First, prior studies have only considered the segmentation of specular and non-specular regions while ignoring the similarity of the content within the specular region. Second, these methods rely on a large number of parameters to achieve good segmentation performance. To address these issues, in this study, we propose a frequency domain fusion network (FDFNet). In addition, to better establish content similarities, we introduce the frequency–time blending module to extract the combined features of the red, green, blue, and depth frequency domains. We also propose a multiscale fusion enhancement module to improve multiscale semantic features. In addition, we combine a semantically complementary comparative learning refinement approach combining comparative learning and knowledge refinement, which divides knowledge into foreground and background. For both teachers and students, the same foreground or background features are brought closer together, while different foreground or background features are moved further apart. The FDFNet-S student model is improved under the guidance of the FDFNet-T teacher model, resulting in FDFNet-S*. Our design compresses the number of parameters and floating-point operands from 164.69 million and 78.86 billion in FDFNet-T to 17.41 million and 7.76 billion in FDFNet-S, respectively. The experimental results confirm the effectiveness of the proposed methodology and framework. The code is available at https://github.com/sfsffafasf/FDFNet.