Scholarlink

<p>Image quality assessment (IQA) has emerged as a crucial research area in computer vision. Most current IQA methods rely heavily on a large amount of labeled training data. However, labeling distorted images is a difficult and time-consuming task. This complicates the acquisition of sufficient training images for specific distortion types. Moreover, diverse application scenarios introduce various complex distortions, which complicate feature extraction and model training. To address these challenges, this paper proposes a two-stream model that effectively integrates local and global feature of the distorted image. In the local feature extraction module, a convolution-based multi-scale feature network is implemented to effectively extract the local feature. Meanwhile, the global context awareness module based on the Vision Transformer (ViT) captures the global information. Furthermore, a meta-learning framework is utilized to enable the model to acquire prior knowledge of various distortion types from a limited number of samples, which allows the model to rapidly adapt to unknown distortion types. This method reduces the dependence on the number of training samples and adapts to different application scenarios. Extensive experimental results demonstrate that the proposed method achieves state-of-the-art (SOTA) performance on public datasets such as LIVEC and KonIQ-10K. Compared to existing IQA methods, this method gains stronger generalization ability and stability. The code is available at <a href="https://github.com/dart-into/MGLIQA">https://github.com/dart-into/MGLIQA</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

Meta-learning enhanced global–local feature fusion for image quality assessment

  • 期刊论文

  • 发表日期：

    2025年5月15日

查看全文
  • Nengxin Li,
  • Xichen Yang,
  • Tianhai Chen,
  • Shun Zhu,
  • Zhongyuan Mao,
  • Tianshu Wang,
  • Xiaobo Shen

摘要

Image quality assessment (IQA) has emerged as a crucial research area in computer vision. Most current IQA methods rely heavily on a large amount of labeled training data. However, labeling distorted images is a difficult and time-consuming task. This complicates the acquisition of sufficient training images for specific distortion types. Moreover, diverse application scenarios introduce various complex distortions, which complicate feature extraction and model training. To address these challenges, this paper proposes a two-stream model that effectively integrates local and global feature of the distorted image. In the local feature extraction module, a convolution-based multi-scale feature network is implemented to effectively extract the local feature. Meanwhile, the global context awareness module based on the Vision Transformer (ViT) captures the global information. Furthermore, a meta-learning framework is utilized to enable the model to acquire prior knowledge of various distortion types from a limited number of samples, which allows the model to rapidly adapt to unknown distortion types. This method reduces the dependence on the number of training samples and adapts to different application scenarios. Extensive experimental results demonstrate that the proposed method achieves state-of-the-art (SOTA) performance on public datasets such as LIVEC and KonIQ-10K. Compared to existing IQA methods, this method gains stronger generalization ability and stability. The code is available at https://github.com/dart-into/MGLIQA.