Scholarlink

<p>The increasing reliance on multimedia applications has brought significant attention to blind image quality assessment (BIQA), as these applications often introduce distortions during stages, such as compression, transmission, and processing. Existing deep-learning-based BIQA approaches face several challenges: processing an entire image can result in the loss of critical fine details due to resizing. At the same time, patch-based sampling may yield patch scores inconsistent with the overall image quality. To address these limitations, the proposed Fine-grained Attention-based Sampling network with Global Features (FEATS-GF) employs selective patch sampling guided by an attention mechanism for local quality assessment, utilizing a patch-order-independent bag-of-features framework. Concurrently, global quality is estimated from the entire image. This end-to-end trainable model synergistically combines global image structures with detailed, attention-driven patch features. Extensive evaluations demonstrate FEATS-GF’s superior performance; for instance, it achieves state-of-the-art Spearman Rank Order Correlation Coefficient (SROCC) and Pearson Linear Correlation Coefficient (PLCC) values of 0.989/0.989 on the LIVE dataset and a notable SROCC of 0.932 on the challenging, authentically distorted KonIQ-10k dataset, significantly outperforming contemporary methods. FEATS-GF thus offers a robust and accurate solution by effectively integrating both macroscopic and microscopic visual information for BIQA.</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

An efficient blind image quality assessment by combining fine-grained attention sampling with global features

  • 期刊论文

  • 发表日期：

    2026年2月27日

查看全文
  • Nandhini Chockalingam,
  • Brindha Murugan,
  • Sharfuddin Waseem Mohammed

摘要

The increasing reliance on multimedia applications has brought significant attention to blind image quality assessment (BIQA), as these applications often introduce distortions during stages, such as compression, transmission, and processing. Existing deep-learning-based BIQA approaches face several challenges: processing an entire image can result in the loss of critical fine details due to resizing. At the same time, patch-based sampling may yield patch scores inconsistent with the overall image quality. To address these limitations, the proposed Fine-grained Attention-based Sampling network with Global Features (FEATS-GF) employs selective patch sampling guided by an attention mechanism for local quality assessment, utilizing a patch-order-independent bag-of-features framework. Concurrently, global quality is estimated from the entire image. This end-to-end trainable model synergistically combines global image structures with detailed, attention-driven patch features. Extensive evaluations demonstrate FEATS-GF’s superior performance; for instance, it achieves state-of-the-art Spearman Rank Order Correlation Coefficient (SROCC) and Pearson Linear Correlation Coefficient (PLCC) values of 0.989/0.989 on the LIVE dataset and a notable SROCC of 0.932 on the challenging, authentically distorted KonIQ-10k dataset, significantly outperforming contemporary methods. FEATS-GF thus offers a robust and accurate solution by effectively integrating both macroscopic and microscopic visual information for BIQA.