Research map / Label-efficient and robust learning / Domain adaptation and generalization
DDJND: Dual Domain Just Noticeable Difference in Multi-Source Content Images with Structural Discrepancy
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes DDJND, a dual-domain (spatial + DCT frequency) encoder-decoder with mutual knowledge distillation that estimates just noticeable difference maps for multi-source content images with structural discrepancies.
Problem
Single-domain JND estimators struggle with structural discrepancies in multi-source content (AIGC, screen content, UGC) images, limiting visual redundancy estimation.
Approach
A dual-domain encoder combines spatial pattern balance (global/local pattern balancing with attention) and frequency detail correction (DCT-residual low/high frequency decomposition and correction), trained with alternating mutual knowledge distillation between the two domains; a dual-domain decoder uses KAN-based multi-scale activation and multi-domain integration.
Key results
- On SHEN2020, DDJND reaches SSIM 0.9283 and LPIPS 0.2212 versus the best baseline Jiang2024 at 0.9251 SSIM.
- Ablation on SHEN2020: the full model (SPI+FDE+MA+MI) gives SSIM 0.9283 and LPIPS 0.2212, against 0.8583 SSIM and 0.3296 LPIPS with all modules disabled.
- On cross-dataset MSC evaluation at MSE=100, DDJND reports SSIM 0.8715 (AIGC), 0.8432 (SC), 0.8206 (UGC), exceeding Jiang2024 on all three.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DDJND instead of Jiang2024 | JND prediction · MCL-JCI | SSIM: 0.9020 -> 0.9124 | Helped |
| Use DDJND instead of Jiang2022 | JND prediction · MCL-JCI | LPIPS: 0.2300 -> 0.2045 | Helped |
| Use DDJND instead of Jiang2024 | JND prediction · SHEN2020 | SSIM: 0.9251 -> 0.9283 | Helped |
| Use DDJND instead of Jiang2024 | JND prediction · SHEN2020 | LPIPS: 0.1043 -> 0.2212 | Hurt |
| Use DDJND instead of Jiang2024 | JND prediction · KonJND1k | SSIM: 0.9409 -> 0.9492 | Helped |
| Use DDJND instead of Jiang2022 | JND prediction · KonJND1k | LPIPS: 0.1501 -> 0.1452 | Helped |
| Use DDJND instead of Jiang2024 | JND prediction · AIGC | SSIM: 0.8648 -> 0.8715 | Helped |
| Use DDJND instead of Jiang2024 | JND prediction · AIGC | LPIPS: 0.1941 -> 0.1883 | Helped |
| Use DDJND instead of Jiang2024 | JND prediction · SC | SSIM: 0.8305 -> 0.8432 | Helped |
| Use DDJND instead of Jiang2024 | JND prediction · SC | LPIPS: 0.2504 -> 0.2352 | Helped |
| Use DDJND instead of Jiang2024 | JND prediction · UGC | SSIM: 0.8076 -> 0.8206 | Helped |
| Use DDJND instead of Jiang2024 | JND prediction · UGC | LPIPS: 0.3964 -> 0.3753 | Helped |
Limitations
- Evaluation relies on SSIM and LPIPS against injected-noise JND proxies rather than subjective tests.
- MSC datasets lack ground-truth JND maps, so they are evaluated at a fixed injected MSE=100 level.
- Training is reported only on SHEN2020 with a small batch size and 40 epochs.
Experimental setup
Datasets: SHEN2020 MCL-JCI KonJND1k DiffusionDB SCID SIQAD Youtube UGC
Benchmarks: MCL-JCI SHEN2020 KonJND1k AIGC (DiffusionDB) SC (SCID, SIQAD) UGC (Youtube UGC)
Metrics: SSIM LPIPS
Baselines: Wei2009 Wu2013 Ki2018 Shen2020 Jiang2022 Jiang2024
Builds on: Wu2017 pattern complexity JND Shen2020 patch-level JND Jiang2024 JND as difference map mutual knowledge distillation
Closest papers by meaning
- D-GAP: Improving Out-of-Domain Robustness via Dataset-Agnostic and Gradient-Guided Augmentation in Amplitude and Pixel Spaces (NeurIPS 2026 · similarity 0.83)
- Misalignment-Robust Frequency Distribution Loss for Image Transformation (CVPR 2024 · similarity 0.82)
- Visual Redundancy Removal for Composite Images: A Benchmark Dataset and a Multi-Visual-Effects Driven Incremental Method (AAAI 2024 · similarity 0.82)
- DCDepth: Progressive Monocular Depth Estimation in Discrete Cosine Domain (NeurIPS 2024 · similarity 0.82)
- Reconstruction Target Matters in Masked Image Modeling for Cross-Domain Few-Shot Learning (AAAI 2025 · similarity 0.82)
- Beyond Spatial Domain: Cross-domain Promoted Fourier Convolution Helps Single Image Dehazing (AAAI 2025 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.