Research map / Vision-language and multimodal / Human and embodied modeling
Visual Redundancy Removal for Composite Images: A Benchmark Dataset and a Multi-Visual-Effects Driven Incremental Method
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.
Builds ciVRP-Set, a 413-image composite-image visual redundancy prediction benchmark, and proposes mveVRP, a multi-visual-effect incremental learning encoder-decoder (IL-VRPnet) for visual redundancy prediction that guides VVC compression.
Problem
Visual redundancy prediction (VRP) for composite images (CIs) is hard because CI content is complex, existing knowledge-driven methods fail to model interactions among visual effects, data-driven methods suffer from label bias, and no benchmark CI VRP dataset exists.
Approach
Combine hand-crafted visual effect knowledge with deep learning by generating five visual-effect data subsets (luminance adaptation, contrast masking, blur sensitivity, orientation masking, visual preference sensitivity), training an encoder-decoder network incrementally in three stages with relative-entropy-based correlation ratios and parameter-freezing regularization to avoid catastrophic forge
Key results
- On ciVRP-Set, the proposed method reaches PSNR 30.6433 and SSIM 0.9359, the best of the eight compared methods in that column (Xie23AAAI: 30.8625 / 0.9243).
- Average across three datasets: PSNR 30.3552 and SSIM 0.9319 versus 30.5203 / 0.9223 for the next best method (Xie23AAAI).
- VVC-guided compression achieves a bit-rate saving of -35.08% versus lossless VVC, the highest among the eight VRPs (Xie23AAAI reports -32.71%, the next best).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use the proposed method instead of Xie23AAAI | visual redundancy prediction · MCI-JCL | PSNR: 30.7109 -> 30.4911 | Helped |
| use the proposed method instead of Jiang22TIP | visual redundancy prediction · MCI-JCL | SSIM: 0.9162 -> 0.9196 | Helped |
| use the proposed method instead of Xie23AAAI | visual redundancy prediction · Shen2020 | PSNR: 29.9875 -> 29.9312 | Helped |
| use the proposed method instead of Xie23AAAI | visual redundancy prediction · Shen2020 | SSIM: 0.9377 -> 0.9403 | Helped |
| use the proposed method instead of Xie23AAAI | visual redundancy prediction · ciVRP-Set | PSNR: 30.8625 -> 30.6433 | Helped |
| use the proposed method instead of Jiang22TIP | visual redundancy prediction · ciVRP-Set | SSIM: 0.9296 -> 0.9359 | Helped |
| add the three-stage visual-effect incremental training scheme instead of direct training (ablation) | IL-VRPnet · ciVRP-Set | SSIM: 0.9243 -> 0.9359 | Helped |
| add the three-stage visual-effect incremental training scheme instead of separate training (ablation) | IL-VRPnet · ciVRP-Set | SSIM: 0.9240 -> 0.9359 | Helped |
| add Stage II regulation to sequential MVE-subset training (ablation) | IL-VRPnet · ciVRP-Set | SSIM: 0.9266 -> 0.9359 | Helped |
| replace the U-Net-like backbone with IL-VRPnet (ablation) | visual redundancy prediction · ciVRP-Set | SSIM: 0.9226 -> 0.9359 | Helped |
| replace the VGG-like encoder with the ResNet-18 encoder (ablation) | IL-VRPnet · ciVRP-Set | SSIM: 0.9340 -> 0.9359 | Helped |
Surprising findings
- Most prior HVS-based methods performed worse on ciVRP-Set than on the natural image datasets, suggesting CI complexity hurts traditional methods.
- Naive consecutive training across visual-effect subsets without regularization yields worse SSIM than the proposed staged scheme, showing the incremental constraint matters.
Limitations
- Subjective evaluation showed over-estimation of redundancy in facial regions, attributed to missing facial knowledge in incremental training.
- Only one GPU type is reported and the dataset is limited to 413 images at a single resolution (1920x1080).
Experimental setup
Datasets: ciVRP-Set MCL-JCI SHEN2020
Benchmarks: ciVRP-Set MCL-JCI SHEN2020
Metrics: PSNR SSIM MOS bit-rate saving bpp
Baselines: Wang16TIP Wu17TIP Chen19TCSVT Shen20TIP Wang22TII Jiang22TIP Xie23AAAI Unet VGG
Builds on: ResNet U-Net VVC
Closest papers by meaning
- Every SAM Drop Counts: Embracing Semantic Priors for Multi-Modality Image Fusion and Beyond (CVPR 2025 · similarity 0.83)
- R4-CGQA: Retrieval-based Vision Language Models for Computer Graphics Image Quality Assessment (CVPR 2026 · similarity 0.83)
- Novel View Synthesis with View-Dependent Effects from a Single Image (CVPR 2024 · similarity 0.83)
- VeXKD: The Versatile Integration of Cross-Modal Fusion and Knowledge Distillation for 3D Perception (NeurIPS 2024 · similarity 0.83)
- MVSplat360: Feed-Forward 360 Scene Synthesis from Sparse Views (NeurIPS 2024 · similarity 0.83)
- RadOcc: Learning Cross-Modality Occupancy Knowledge through Rendering Assisted Distillation (AAAI 2024 · similarity 0.83)
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.