Research map / Image generation / Text-to-image diffusion
Hierarchical Adaptive Filtering Network for Text Image Specular Highlight Removal
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 HAFNet, a kernel-prediction filtering network that removes large-area specular highlights from text images by filtering at both a deep feature layer (ACFM) and an image reconstruction layer (ADFM), and releases a Unity3D-rendered large-area highlight dataset (LSH).
Problem
Removing large, intense specular highlights from text images (cards, posters) where flat surfaces produce large highlight areas, causing artifacts and detail loss that existing methods, trained mostly on small natural-image highlights, fail to handle.
Approach
Hierarchical kernel-prediction filtering: a large predicted kernel (15x15) applied in the frequency domain via FFT with separable outer-product kernels (POFM) at the downsampled deep feature layer for semantic restoration, plus multi-scale dilated 3x3 kernels at the reconstruction layer for detail refinement.
Key results
- On the LSH dataset HAFNet reaches 26.593 PSNR and 0.937 SSIM, versus 25.066 PSNR / 0.905 SSIM for the next-best TSHR.
- On SD1, HAFNet reports 28.952 PSNR vs 25.147 for TSHR; on RD it reports 26.217 PSNR vs 23.903 for IDRHR.
- Ablation on LSH: removing ACFM drops PSNR to 22.017 and removing ADFM to 25.711, showing both modules contribute.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add ACFM deep-feature filtering (ablation) | HAFNet · LSH | PSNR: 22.017 -> 26.593 | Helped |
| Add ACFM deep-feature filtering (ablation) | HAFNet · LSH | SSIM: 0.849 -> 0.937 | Helped |
| Add ADFM image-reconstruction-layer filtering (ablation) | HAFNet · LSH | PSNR: 25.711 -> 26.593 | Helped |
| Add ADFM image-reconstruction-layer filtering (ablation) | HAFNet · LSH | SSIM: 0.907 -> 0.937 | Helped |
| Use ACFM instead of standard KPF (ablation) | HAFNet · LSH | PSNR: 25.148 -> 26.593 | Helped |
| Use ACFM instead of standard KPF (ablation) | HAFNet · LSH | SSIM: 0.912 -> 0.937 | Helped |
| Use ADFM instead of standard KPF (ablation) | HAFNet · LSH | PSNR: 26.019 -> 26.593 | Helped |
| Use ADFM instead of standard KPF (ablation) | HAFNet · LSH | SSIM: 0.921 -> 0.937 | Helped |
| Set ACFM large-kernel size to k=11, k=13, k=17, or k=19 instead of k=15 (ablation) | HAFNet · LSH | PSNR: 25.698; 26.284; 26.603; 26.611 (k=11; k=13; k=17; k=19), versus 26.593 (k=15) | Mixed |
| Set ACFM large-kernel size to k=11, k=13, k=17, or k=19 instead of k=15 (ablation) | HAFNet · LSH | SSIM: 0.929; 0.931; 0.933; 0.932 (k=11; k=13; k=17; k=19), versus 0.937 (k=15) | Mixed |
| Use i=1, i=2, or i=4 predicted kernels in ADFM instead of i=3 (ablation) | HAFNet · LSH | PSNR: 26.019; 26.532; 26.607 (i=1; i=2; i=4), versus 26.593 (i=3) | Mixed |
| Use i=1, i=2, or i=4 predicted kernels in ADFM instead of i=3 (ablation) | HAFNet · LSH | SSIM: 0.921; 0.929; 0.930 (i=1; i=2; i=4), versus 0.937 (i=3) | Mixed |
Limitations
- Performance degrades when highlight areas are very large and very intense, lacking reliable information.
- Restoring details under complex highlights remains difficult, often producing blur.
- Evaluation on real highlight images is only qualitative due to missing ground truth.
Experimental setup
Datasets: LSH (proposed, Unity3D-rendered) RD SD1
Benchmarks: LSH RD SD1
Metrics: PSNR SSIM
Baselines: Shen Yang TASHR JSHDR Wu TSHR IDRHR
Builds on: kernel prediction filtering CBAM relativistic average GAN dilated convolution
Closest papers by meaning
- SHDocs: A dataset, benchmark, and method to efficiently generate high-quality, real-world specular highlight data with near-perfect alignment (NeurIPS 2024 · similarity 0.86)
- PHR-DIFF: Portrait Highlights Removal via Patch-aware Diffusion Model (AAAI 2025 · similarity 0.85)
- FIRM: Flexible Interactive Reflection ReMoval (AAAI 2025 · similarity 0.82)
- Specularity Factorization for Low-Light Enhancement (CVPR 2024 · similarity 0.82)
- AMO Sampler: Enhancing Text Rendering with Overshooting (CVPR 2025 · similarity 0.81)
- Seeing What Matters: Empowering CLIP with Patch Generation-to-Selection (CVPR 2025 · similarity 0.81)
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.