GhostSR: Learning Ghost Features for Efficient Image Super-Resolution
Authors: Ying Nie, Kai Han, Zhenhua Liu, Chuanjian Liu, Yunhe Wang
TMLR 2022 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on several benchmark models and datasets demonstrate that both the non-compact and lightweight SISR CNN models embedded with the proposed method can achieve a comparable performance to the baseline models with a large reduction of parameters, FLOPs and GPU inference latency. |
| Researcher Affiliation | Industry | Ying Nie EMAIL Huawei Noah s Ark Lab Kai Han EMAIL Huawei Noah s Ark Lab Zhenhua Liu EMAIL Huawei Noah s Ark Lab Chuanjian Liu EMAIL Huawei Noah s Ark Lab Yunhe Wang EMAIL Huawei Noah s Ark Lab |
| Pseudocode | No | The paper describes the method using mathematical equations and descriptive text, and Figure 3 provides an example of the shift operation, but there is no explicitly labeled pseudocode or algorithm block with structured steps. |
| Open Source Code | No | Code will be available at https://gitee.com/mindspore/models/tree/master/research/cv/Ghost SR. |
| Open Datasets | Yes | To evaluate the performance of the method, following the setting of (Lim et al., 2017; Zhang et al., 2018c; Ahn et al., 2018), we use 800 images from DIV2K (Timofte et al., 2017) dataset to train our models. In order to compare with other state-of-the-art methods, we report our result on four standard benchmark datasets: Set5 (Bevilacqua et al., 2012), Set14 (Zeyde et al., 2010), B100 (Martin et al., 2001) and Urban100 (Huang et al., 2015). |
| Dataset Splits | Yes | To evaluate the performance of the method, following the setting of (Lim et al., 2017; Zhang et al., 2018c; Ahn et al., 2018), we use 800 images from DIV2K (Timofte et al., 2017) dataset to train our models. In order to compare with other state-of-the-art methods, we report our result on four standard benchmark datasets: Set5 (Bevilacqua et al., 2012), Set14 (Zeyde et al., 2010), B100 (Martin et al., 2001) and Urban100 (Huang et al., 2015). |
| Hardware Specification | Yes | Table 2: Average inference latency (ms) for Urban100 dataset with 2 scale on a single V100 GPU platform. |
| Software Dependencies | No | The paper mentions 'Mind Spore (Huawei, 2020)', 'ADAM optimizer (Kingma & Ba, 2014)', and 'cosine learning rate decay' but does not specify version numbers for any software libraries or frameworks. 'Huawei, 2020' refers to a publication date for MindSpore, not its version. |
| Experiment Setup | Yes | During training, we crop 16 images with 48 48 patch size from the LR images on every card for training. The input examples are augmented by random horizontal flipping and 90 rotating. In addition, all the images are pre-processed by subtracting the mean RGB value of the DIV2K dataset. To optimize the model, we use ADAM optimizer (Kingma & Ba, 2014) with β1 = 0.9, β2 = 0.999, and ϵ = 10 8. We train EDSR and RDN for 300 epochs by single-scale training scheme, and train CARN and IMDN for 1200 epochs by multi-scale and single-scale training scheme respectively. The initial learning rate is set to 1e-4 for all models and reduced by cosine learning rate decay (Zhao et al., 2020; Kong et al., 2021). Unless otherwise specified, the ratio λ of ghost features is set to 0.5, the temperature τ in Eq. 4 is set to 1, and the maximum offset d in Eq. 1 is set to 1. |