Scholarlink

<p>Image Quality Assessment (IQA) is crucial in image processing and visual optimization. However, real-world image distortions are complex and varied, which makes it essential to develop models specifically trained to handle authentic distortions. Effective training requires high-quality labeled data for supervised learning. Unfortunately, obtaining accurate quality labels is challenging due to the subjective nature of image quality and the high cost of manual annotation, limiting model training and generalization. To address this challenge, this paper proposes a cross-domain image quality assessment method based on Adaptive Similarity Domain Selection (ASDS-IQA), aiming to adapt existing data to new distortion types. Firstly, this paper proposes a distortion classification module based on adaptive thresholds by combining mean and standard deviation, and smooths the classification probability distribution through a temperature scaling mechanism, thereby enhancing the model’s adaptability to complex distortion types. Secondly, this paper selects data from the source domain that exhibits distribution similarity to the target domain for model training, reducing the negative transfer effect caused by inter-domain differences and further improving the model’s generalization performance on the target domain. Finally, a multi-level feature fusion module is proposed, which achieves efficient and stable performance in cross-domain scenarios by effectively integrating feature information from different convolutional layers. Experimental results demonstrate that the proposed method achieves competitive performance on public datasets such as LIVEC and KonIQ-10K. Compared to existing image quality assessment methods, the proposed method exhibits superior accuracy and generalization capabilities. The code is available at <a href="https://github.com/dart-into/ASDS-IQA">https://github.com/dart-into/ASDS-IQA</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

Cross-domain image quality assessment method based on adaptive similarity domain selection

  • 期刊论文

  • 发表日期：

    2026年5月15日

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

摘要

Image Quality Assessment (IQA) is crucial in image processing and visual optimization. However, real-world image distortions are complex and varied, which makes it essential to develop models specifically trained to handle authentic distortions. Effective training requires high-quality labeled data for supervised learning. Unfortunately, obtaining accurate quality labels is challenging due to the subjective nature of image quality and the high cost of manual annotation, limiting model training and generalization. To address this challenge, this paper proposes a cross-domain image quality assessment method based on Adaptive Similarity Domain Selection (ASDS-IQA), aiming to adapt existing data to new distortion types. Firstly, this paper proposes a distortion classification module based on adaptive thresholds by combining mean and standard deviation, and smooths the classification probability distribution through a temperature scaling mechanism, thereby enhancing the model’s adaptability to complex distortion types. Secondly, this paper selects data from the source domain that exhibits distribution similarity to the target domain for model training, reducing the negative transfer effect caused by inter-domain differences and further improving the model’s generalization performance on the target domain. Finally, a multi-level feature fusion module is proposed, which achieves efficient and stable performance in cross-domain scenarios by effectively integrating feature information from different convolutional layers. Experimental results demonstrate that the proposed method achieves competitive performance on public datasets such as LIVEC and KonIQ-10K. Compared to existing image quality assessment methods, the proposed method exhibits superior accuracy and generalization capabilities. The code is available at https://github.com/dart-into/ASDS-IQA.