Research map / Detection and segmentation / CNN features
Wavelet and Prototype Augmented Query-based Transformer for Pixel-level Surface Defect Detection
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 WPFormer, a query-based transformer with wavelet-enhanced and prototype-guided cross-attention for pixel-level surface defect mask prediction, evaluated on ESDIs-SOD, CrackSeg9k and ZJU-Leaper.
Problem
Pixel-level surface defect detection struggles with weak defects and cluttered backgrounds, since static convolutional prediction heads are image-independent and query-based decoders lack frequency-domain cues and waste attention on background.
Approach
A dual-domain transformer decoder (D2T) that refines dynamic queries via Haar wavelet-domain cross-attention (WCA, modulating high-frequency subbands with global/local channel weights) and spatial prototype-guided cross-attention (PCA, learnable prototypes obtained by soft clustering of image tokens).
Key results
- On ESDIs-SOD, WPFormer reaches weighted F-measure 0.8901 and MAE 0.0171, versus 0.8863 and 0.0183 for A3Net.
- Averaged over three datasets, WPFormer improves MAE by 12.14%, weighted F-measure by 2.88% and S-measure by 1.21% relative to Mask2Former.
- The combined WCA and PCA cross-attention gives averaged gains of 2.14% weighted F-measure over standard cross-attention on the ablation datasets.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace PEM with WPFormer | PVTv2-based segmentation on ESDIs-SOD · ESDIs-SOD | Fᵝw: .8747 -> .8901 | Helped |
| replace IdeNet with WPFormer | surface defect segmentation · CrackSeg9k | Fᵝw: .7510 -> .7672 | Helped |
| replace ZoomNeXt with WPFormer | surface defect segmentation · ZJU-Leaper | Fᵝw: .7803 -> .7972 | Helped |
| add WCA and PCA instead of conventional cross-attention (ablation) | WPFormer · ESDIs-SOD | Fᵝw: .8778 -> .8901 | Helped |
| add WCA and PCA instead of conventional cross-attention (ablation) | WPFormer · CrackSeg9k | Fᵝw: .7458 -> .7672 | Helped |
| add WCA and PCA instead of masked cross-attention (ablation) | WPFormer · ESDIs-SOD | Fᵝw: .8797 -> .8901 | Helped |
| add WCA and PCA instead of PEM-CA (ablation) | WPFormer · CrackSeg9k | Fᵝw: .7513 -> .7672 | Helped |
| add PCA to WCA (ablation) | WPFormer · ESDIs-SOD | Fᵝw: .8858 -> .8901 | Helped |
| add WCA to PCA (ablation) | WPFormer · CrackSeg9k | Fᵝw: .7579 -> .7672 | Helped |
| use 16 rather than 8 queries (ablation) | WPFormer · ESDIs-SOD | Fᵝw: .8853 -> .8901 | Helped |
| use 16 rather than 64 queries (ablation) | WPFormer · CrackSeg9k | Fᵝw: .7654 -> .7672 | Helped |
| use modulated instead of additive frequency fusion (ablation) | WCA · ESDIs-SOD | Fᵝw: .8865 -> .8901 | Helped |
Limitations
- Only evaluated at 384x384 input resolution on three datasets
- Gains over strong baselines such as IdeNet and MSCAFNet are small (often below 1-2%)
- Relies on fixed Haar wavelet filters that the authors note lack semantics
Experimental setup
Datasets: ESDIs-SOD CrackSeg9k ZJU-Leaper
Benchmarks: ESDIs-SOD CrackSeg9k ZJU-Leaper
Metrics: MAE weighted F-measure S-measure mean F-measure mean E-measure
Baselines: JTFN SINetV2 Mask2Former BBRF PUENet FPNet MENet FSPNet FEDER MSCAFNet A3Net IdeNet
Builds on: Mask2Former PEM MaskFormer PVTv2 FPN
Closest papers by meaning
- MVREC: A General Few-shot Defect Classification Model Using Multi-View Region-Context (AAAI 2025 · similarity 0.85)
- Exploring the Relationship Between Samples and Masks for Robust Defect Localization (AAAI 2025 · similarity 0.84)
- DefectFill: Realistic Defect Generation with Inpainting Diffusion Model for Visual Inspection (CVPR 2025 · similarity 0.84)
- Towards Open-Vocabulary Industrial Defect Understanding with a Large-Scale Multimodal Dataset (CVPR 2026 · similarity 0.82)
- Automated Defect Report Generation for Enhanced Industrial Quality Control (AAAI 2024 · similarity 0.82)
- Few Shot Part Segmentation Reveals Compositional Logic for Industrial Anomaly Detection (AAAI 2024 · 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.