Research map / Image generation / Autoregressive and latent generation

CutFreq: Cut-and-Swap Frequency Components for Low-Level Vision Augmentation

AAAI 2024 · Image generation › Autoregressive and latent generation · approach: Learning-based image quality assessment · 6 citations (OpenAlex)

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 CutFreq, a single-sample augmentation that cuts input and ground-truth images into DWT frequency bands and swaps them in two training stages, improving five low-level vision tasks.

Problem

Standard high-level data augmentations (CutMix, Mixup, Cutout) and existing low-level ones (CutBlur, Copy-blend) break spatial consistency or introduce unrealistic content, giving limited gains on low-level vision tasks with small datasets.

Approach

Decompose input and real images with Haar discrete wavelet transform and swap selected frequency bands between them in two stages (fast convergence, then performance approximation with random band swap or NULL discarding), guided by a frequency-band normalized L1 distance analysis of convergence per band.

Key results

• On Zurich smart ISP with LW-ISP, CutFreq reaches 21.54 dB, +0.17 dB over the 21.37 dB baseline and above all compared augmentations.
• On SIDD denoising with LW-ISP, CutFreq improves PSNR from 39.1966 to 39.2904 dB (+0.094 dB).
• CutFreq gives about 0.11 dB average gain over Restormer on deblurring and 0.06 dB average gain on deraining, and gains persist when training data is reduced to 5%.

Measured effects reported in the paper

Surprising findings

• Common high-level augmentations such as CutBlur and Copy-blend degrade PSNR substantially on smart ISP (about -1.50 dB), and Blend degrades denoising by -1.93 dB.
• Replacing DWT with Daubechies wavelets keeps performance below baseline for J1 and J2, while Haar consistently helps.

Limitations

• Effectiveness may not be universally optimal because networks do not explicitly learn frequency information.
• Spectral losses (FFL, WL) did not give consistent gains compared with CutFreq.
• Hyperparameters such as decomposition level, wavelet basis, and swap schedule affect performance and were tuned per task.

Experimental setup

Datasets: Zurich RAW to RGB SIDD Rain100H Rain100L Test100 Test2800 Test1200 GoPro HIDE LoL

Benchmarks: Zurich RAW to RGB SIDD DND GoPro HIDE Rain100H Rain100L Test100 Test2800 Test1200 LoL

Metrics: PSNR SSIM

Baselines: Mixup CutMix CutMixup Cutout CutBlur Copy-blend Blend RGB permutation

Builds on: CutMix CutBlur LW-ISP Restormer HINet

Closest papers by meaning

• Fourier-basis Functions to Bridge Augmentation Gap: Rethinking Frequency Augmentation in Image Classification (CVPR 2024 · similarity 0.87)
• Provable Benefit of Cutout and CutMix for Feature Learning (NeurIPS 2024 · similarity 0.86)
• D-GAP: Improving Out-of-Domain Robustness via Dataset-Agnostic and Gradient-Guided Augmentation in Amplitude and Pixel Spaces (NeurIPS 2026 · similarity 0.85)
• Misalignment-Robust Frequency Distribution Loss for Image Transformation (CVPR 2024 · similarity 0.84)
• Frequency-Adaptive Dilated Convolution for Semantic Segmentation (CVPR 2024 · similarity 0.83)
• MimiQ: Low-Bit Data-Free Quantization of Vision Transformers with Encouraging Inter-Head Attention Similarity (AAAI 2025 · similarity 0.83)