Scholarlink

<p>With the rapid development of 3D technology, effective no-reference stereoscopic image quality assessment (NR-SIQA) methods have become increasingly important. This paper proposes a selective feedback-guided stereoscopic image quality assessment network (SFSIQA-Net), which fully considers the visual feedback mechanisms in the human visual system (HVS) and selectively feeds higher-level features back to lower-level features. Additionally, this paper introduces the transposed attention-based binocular fusion module (TBFM), which incorporates a self-attention mechanism to capture global features. By introducing a visual feedback mechanism, it accurately simulates how binocular fusion information guides monocular information, and by adopting a transposed attention block, it enhances the representation of semantic information. Finally, this paper proposes the semantic feature enhancement module (SFEM) and texture feature enhancement module (TFEM), which comprehensively enhance the semantic information of lower-level features and the texture information of higher-level features, facilitating the alignment of features across different levels. The proposed model achieves excellent results on the LIVE I, LIVE II, WIVC I, and WIVC II databases, validating the effectiveness of the proposed model.</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

SFSIQA-Net: stereo image quality assessment based on selective feedback and transposed attention binocular fusion

  • 期刊论文

  • 发表日期：

    2026年7月20日

查看全文
  • Sumei Li,
  • Mingxuan Xie

摘要

With the rapid development of 3D technology, effective no-reference stereoscopic image quality assessment (NR-SIQA) methods have become increasingly important. This paper proposes a selective feedback-guided stereoscopic image quality assessment network (SFSIQA-Net), which fully considers the visual feedback mechanisms in the human visual system (HVS) and selectively feeds higher-level features back to lower-level features. Additionally, this paper introduces the transposed attention-based binocular fusion module (TBFM), which incorporates a self-attention mechanism to capture global features. By introducing a visual feedback mechanism, it accurately simulates how binocular fusion information guides monocular information, and by adopting a transposed attention block, it enhances the representation of semantic information. Finally, this paper proposes the semantic feature enhancement module (SFEM) and texture feature enhancement module (TFEM), which comprehensively enhance the semantic information of lower-level features and the texture information of higher-level features, facilitating the alignment of features across different levels. The proposed model achieves excellent results on the LIVE I, LIVE II, WIVC I, and WIVC II databases, validating the effectiveness of the proposed model.