Research map / Label-efficient and robust learning / Recognition and long-tail
Frequency Shuffling and Enhancement for Open Set Recognition
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 FreSH, which uses DWT-based high-frequency shuffling augmentation and low-frequency residual skip connections to reduce spatial-domain inductive bias and improve open-set recognition on VGG32 benchmarks.
Problem
Open-set recognition suffers from over-confidence on unknown and known classes because spatial-domain CNNs over-perceive subtle high-frequency texture, entangled with global structure.
Approach
Decompose images with Haar DWT, shuffle high-frequency subbands across augmented views while keeping the low-frequency band fixed (HFS), and add skip connections that feed low-frequency subbands from shallow to deep stages (LFR) to form a composite feature space, without extra parameters.
Key results
- FreSH reaches 98.1 AUROC on SVHN and 83.8 AUROC on Tiny-ImageNet, the best among compared methods in Table 1 (except MNIST, where ARPL is 0.1 higher).
- FreSH achieves 87.6% closed-set accuracy on Tiny-ImageNet, a 3.1 point gain over the prior best PMAL (Table 2).
- FreSH improves OSCR on Tiny-ImageNet to 77.9, about 10.1 points above AKPF according to the authors' text.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FreSH instead of the strongest listed baseline | Open-set recognition · SVHN | AUROC: 96.7 -> 98.1 | Helped |
| Use FreSH instead of the strongest listed baseline | Open-set recognition · CIFAR10 | AUROC: 95.0 -> 95.2 | Helped |
| Use FreSH instead of the strongest listed baseline | Open-set recognition · Tiny-ImageNet | OSCR: 67.8 ± NR -> 77.9 ± 4.1 | Helped |
| Add HFS and LFR to the plain CNN (ablation) | CIFAR100 open-set classification · Openness 18.4%–48.9% | Macro F1-Score | Helped |
| Add HFS to the plain CNN (ablation) | CIFAR100 open-set classification · Openness 18.4%–48.9% | Macro F1-Score | Helped |
| Add LFR to the plain CNN (ablation) | CIFAR100 open-set classification · Openness 18.4%–48.9% | Macro F1-Score | Helped |
| Replace DWT with DFT-12 (ablation) | Frequency-based recognition model · CIFAR100 open-set split | AUROC: 83.8 -> 83.6 | Hurt |
| Replace DWT with DFT-16 (ablation) | Frequency-based recognition model · CIFAR100 open-set split | AUROC: 83.8 -> 83.5 | Hurt |
| Replace DWT with DCT-12 (ablation) | Frequency-based recognition model · CIFAR100 open-set split | AUROC: 83.8 -> 83.2 | Hurt |
| Replace DWT with DCT-16 (ablation) | Frequency-based recognition model · CIFAR100 open-set split | AUROC: 83.8 -> 83.0 | Hurt |
| Replace HFS with Gaussian blur (ablation) | FreSH · CIFAR+50 | OSCR: 94.4 -> 91.8 | Hurt |
| Replace HFS with low-frequency-only images (ablation) | FreSH · CIFAR+50 | OSCR: 94.4 -> 83.7 | Hurt |
Surprising findings
- Authors report that CNNs trained on original spatial images show performance trends similar to those trained on low-frequency bands, and that high-frequency damaged inputs can reverse predictions.
Limitations
- Evaluation is limited to small-resolution benchmarks with a VGG32 backbone; generality to larger architectures is only partly tested via fine-grained experiments.
- Gains on MNIST are not state-of-the-art (ARPL is 0.1 higher).
Experimental setup
Datasets: MNIST SVHN CIFAR10 CIFAR+10 CIFAR+50 Tiny-ImageNet CIFAR100 ImageNet LSUN
Benchmarks: MNIST SVHN CIFAR10 CIFAR+10 CIFAR+50 Tiny-ImageNet
Metrics: AUROC closed-set accuracy OSCR macro F1-score
Baselines: Softmax OpenHybrid PROSER ARPL ARPL+CS DIAS BCR ODL KPF PMAL GCPL RPL
Builds on: ARPL PROSER Discrete Wavelet Transform ResNet residual learning
Closest papers by meaning
- Exploring Diverse Representations for Open Set Recognition (AAAI 2024 · similarity 0.88)
- All Beings Are Equal in Open Set Recognition (AAAI 2024 · similarity 0.87)
- From Coarse to Fine-Grained Open-Set Recognition (CVPR 2024 · similarity 0.87)
- Unlocking the Power of Open Set: A New Perspective for Open-Set Noisy Label Learning (AAAI 2024 · similarity 0.85)
- Improving Open Set Recognition via Visual Prompts Distilled from Common-Sense Knowledge (AAAI 2024 · similarity 0.84)
- Unlocking Better Closed-Set Alignment Based on Neural Collapse for Open-Set Recognition (AAAI 2025 · similarity 0.84)
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.