Scholarlink

<p>Vision Transformers (ViT) have made significant progress in computer vision tasks, especially after the introduction of masking operations, which have demonstrated stronger performance in areas such as image classification and detection. However, in self-supervised learning, masking tends to weaken the correlation of local features of an image, which affects the effectiveness of downstream tasks. To address this problem, this paper proposes a new model method, MCDINO, which extends the capability of image mask operation based on the DINO model (based on improved denoising anchor frame DETR). MCDINO introduces a multipath channel attention mechanism to dynamically compute the importance weight of the feature region to capture the key information among local features effectively; meanwhile, it designs a local feature weighted fusion operation in the Block to improve the fineness of local feature expression and correlation modeling capability. Experiments demonstrate that MCDINO significantly outperforms DINO in the ImageNet100 fine-tuning task, and also achieves better performance in downstream tasks such as copy detection. The results show that MCDINO can effectively enhance the feature learning and representation ability under masking conditions and drive the performance of visual tasks. The code has been open-sourced at: <a href="https://github.com/wangzy2024/MCDINO">https://github.com/wangzy2024/MCDINO</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

MCDINO: Self-supervised learning of masks based on combination of multi-path channel attention and local feature weighting

  • 期刊论文

  • 发表日期：

    2025年6月26日

查看全文
  • Yunxue Shao,
  • Zhiyang Wang,
  • Lingfeng Wang

摘要

Vision Transformers (ViT) have made significant progress in computer vision tasks, especially after the introduction of masking operations, which have demonstrated stronger performance in areas such as image classification and detection. However, in self-supervised learning, masking tends to weaken the correlation of local features of an image, which affects the effectiveness of downstream tasks. To address this problem, this paper proposes a new model method, MCDINO, which extends the capability of image mask operation based on the DINO model (based on improved denoising anchor frame DETR). MCDINO introduces a multipath channel attention mechanism to dynamically compute the importance weight of the feature region to capture the key information among local features effectively; meanwhile, it designs a local feature weighted fusion operation in the Block to improve the fineness of local feature expression and correlation modeling capability. Experiments demonstrate that MCDINO significantly outperforms DINO in the ImageNet100 fine-tuning task, and also achieves better performance in downstream tasks such as copy detection. The results show that MCDINO can effectively enhance the feature learning and representation ability under masking conditions and drive the performance of visual tasks. The code has been open-sourced at: https://github.com/wangzy2024/MCDINO