CutFreq: Cut-and-Swap Frequency Components for Low-Level Vision Augmentation
Authors: Hongyang Chen, Kaisheng Ma
AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | The experimental results show the superior performance of Cut Freq on five low-level vision tasks. |
| Researcher Affiliation | Academia | Hongyang Chen1, Kaisheng Ma2* 1Xi an Jiaotong University, Xi an, China 2Tsinghua University, Beijing, China chenhy@stu.xjtu.edu.cn, kaisheng@mail.tsinghua.edu.cn |
| Pseudocode | No | The paper does not contain any section or figure explicitly labeled 'Pseudocode' or 'Algorithm', nor does it present structured steps in a pseudocode-like format. |
| Open Source Code | Yes | Code is available at https://github.com/DreamerCCC/CutFreq. |
| Open Datasets | Yes | We evaluate the proposed Cut Freq on benchmark datasets and experimental settings for five low-level vision tasks: (a) smart ISP, (b) image denoising, (c) image deraining, (d) image deblurring, and (e) image enhancement. For smart ISP, we evaluate our method on Zurich RAW to RGB (Ignatov, Van Gool, and Timofte 2020) (Zurich for short) dataset. ... We train HINet (Chen et al. 2021b) and Restormer (Zamir et al. 2021) on the training set of SIDD (Abdelhamed, Lin, and Brown 2018) and directly evaluate it on the test images of both SIDD and DND (Plotz and Roth 2017) datasets. ... We evaluate image deblurring on Go Pro (Nah, Hyun Kim, and Mu Lee 2017) and HIDE (Shen et al. 2019). ... We also achieve quite competitive results on the low-light Lo L dataset (Wei et al. 2018). |
| Dataset Splits | No | The paper mentions using 'training set of SIDD' and 'test images of both SIDD and DND' but does not explicitly provide specific training/validation/test dataset splits, percentages, or absolute sample counts within the main text. |
| Hardware Specification | Yes | The experiments are implemented with PyTorch 1.2.0 on RTX NVIDIA 2080Ti and PH402 SKU 200 with 12G memory GPUs. |
| Software Dependencies | Yes | The experiments are implemented with PyTorch 1.2.0 on RTX NVIDIA 2080Ti and PH402 SKU 200 with 12G memory GPUs. |
| Experiment Setup | Yes | We evaluate the proposed Cut Freq on benchmark datasets and experimental settings for five low-level vision tasks... Implementation Details. The experiments are implemented with Py Torch 1.2.0 on RTX NVIDIA 2080Ti and PH402 SKU 200 with 12G memory GPUs. In experiments, PSNR (Huynh-Thu and Ghanbari 2008) and SSIM (Wang et al. 2004) are used to evaluate the image quality. ... Ablation Study ... (a) The frequency information of different decomposition times (denoted as Jn) is used as the cutting standard instead of different frequency bands. ... (b) Under the strategy of maintaining the first stage, we replace another frequency band every certain number of training epochs (10 epochs and 20 epochs). ... Hyperparameters. ... (a) Decomposition times. In the image denoising of LW-ISP, for example, Cut Freq with J2 (+0.06 d B) and J4 (+0.03 d B) slightly outperforms J1 (-0.10 d B). |