Scholarlink

<p>The Human Visual System (HVS) cannot perceive image pixel intensity changes below a certain threshold, which refers to the concept of Just-Noticeable Difference (JND). This concept is significant in numerous applications related to HVS and can benefit many perceptual image processing tasks. Recently, data-driven approaches have excelled in the JND domain by learning directly from data. However, annotating large-scale datasets with pristine images and the corresponding pixel-wise JND maps is an extremely difficult task. Inspired by the fact that JND is conceptually defined as the difference between the pristine image and its corresponding Critical Perceptual Lossless (CPL) image, our previous work constructs a CPL image dataset called CPL-Set to facilitate learning an implicit deep JND model from pristine and CPL image pairs. However, the CPL-Set is limited in size since annotating large-scale CPL images is still labor-consuming. To address this issue, we in this paper utilize the well-known distortion visibility metric, i.e., High Dynamic Range Visual Difference Predictor (HDR-VDP), as an automatic tool to efficiently generate pseudo CPL labels, thereby creating a large-scale dataset with 8,&#xa0;965 high-resolution pristine images and their corresponding pseudo CPL images as labels for JND model training. The newly JND dataset is called CPL-Set2. On the basis of CPL-Set2, we further introduce a lightweight deep invertible neural network for JND estimation, i.e., InvJND. The key consideration is that implicit visual redundancy information, closely related with JND, can be well-removed in the latent deep feature space built via the forward process and then the CPL image can be well reconstructed via the backward process. Comparison experiments demonstrate that, even with pseudo CPL labels for model training, our proposed InvJND model still achieves significant predictive accuracy of JND. In addition, InvJND also exhibits outstanding performance in comprehensive perceptual image processing tasks including JND-guided noise injection, JND-guided compression, JND-guided adversarial attack, and distortion visibility detection. Our codes and datasets are available at: <a href="https://github.com/Knife646/InvJND">https://github.com/Knife646/InvJND</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

InvJND: Just Noticeable Difference Estimation via Deep Invertible Network

  • 期刊论文

  • 发表日期：

    2025年12月22日

查看全文
  • Qiuping Jiang,
  • Feiyang Liu,
  • Ziqi Wang,
  • Zhihua Wang,
  • Shiqi Wang,
  • Feng Shao,
  • Guangtao Zhai,
  • Weisi Lin

摘要

The Human Visual System (HVS) cannot perceive image pixel intensity changes below a certain threshold, which refers to the concept of Just-Noticeable Difference (JND). This concept is significant in numerous applications related to HVS and can benefit many perceptual image processing tasks. Recently, data-driven approaches have excelled in the JND domain by learning directly from data. However, annotating large-scale datasets with pristine images and the corresponding pixel-wise JND maps is an extremely difficult task. Inspired by the fact that JND is conceptually defined as the difference between the pristine image and its corresponding Critical Perceptual Lossless (CPL) image, our previous work constructs a CPL image dataset called CPL-Set to facilitate learning an implicit deep JND model from pristine and CPL image pairs. However, the CPL-Set is limited in size since annotating large-scale CPL images is still labor-consuming. To address this issue, we in this paper utilize the well-known distortion visibility metric, i.e., High Dynamic Range Visual Difference Predictor (HDR-VDP), as an automatic tool to efficiently generate pseudo CPL labels, thereby creating a large-scale dataset with 8, 965 high-resolution pristine images and their corresponding pseudo CPL images as labels for JND model training. The newly JND dataset is called CPL-Set2. On the basis of CPL-Set2, we further introduce a lightweight deep invertible neural network for JND estimation, i.e., InvJND. The key consideration is that implicit visual redundancy information, closely related with JND, can be well-removed in the latent deep feature space built via the forward process and then the CPL image can be well reconstructed via the backward process. Comparison experiments demonstrate that, even with pseudo CPL labels for model training, our proposed InvJND model still achieves significant predictive accuracy of JND. In addition, InvJND also exhibits outstanding performance in comprehensive perceptual image processing tasks including JND-guided noise injection, JND-guided compression, JND-guided adversarial attack, and distortion visibility detection. Our codes and datasets are available at: https://github.com/Knife646/InvJND.