C-Face: Using Compare Face on Face Hallucination for Low-Resolution Face Recognition
Authors: Feng Han, Xudong Wang, Furao Shen, Jian Zhao
JAIR 2022 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on two large scale face datasets demonstrate that our C-Face network has the best performance compared with other state-of-the-art methods. |
| Researcher Affiliation | Academia | Feng Han EMAIL Xudong Wang EMAIL State Key Laboratory for Novel Software Technology, Nanjing University, Nanjing 210046, China Department of Computer Science and Technology, Nanjing University, Nanjing 210046, China Furao Shen (corresponding author) EMAIL State Key Laboratory for Novel Software Technology, Nanjing University, Nanjing 210046, China School of Artificial Intelligence, Nanjing University, Nanjing 210046, China Jian Zhao (corresponding author) EMAIL School of Electronic Science and Engineering, Nanjing University, Nanjing 210046, China |
| Pseudocode | Yes | Algorithm 1 Training Procedure of C-Face Network Input: CNNR pretrained by HR face images, the GAN model including NSR and ND, a set of three images including ILR o , IHR o and IHR c . 1: Train NSR and ND with the loss function in (7); 2: Use the NSR that we obtained after stage 1 to process all the LR images. Now, we have two datasets: the original HR dataset DHR and the super-resolved dataset DSR; 3: Mix DHR and DSR together, and use the combined dataset to fine-tune CNNR; 4: Fine-tune NSR and ND by using the CNNR after stage 3 to extract features. In each step, we update NSR and ND by descending the loss function (7), we also update the CNNR with LFR in (3); Output: The final C-Face network. |
| Open Source Code | No | The paper does not provide an explicit statement or link for open-source code for the methodology described. |
| Open Datasets | Yes | In our experiments, we use the CASIA-Web Face dataset (Yi et al., 2014) with 494, 414 images as our training set and use the LFW dataset (Huang et al., 2008) with 13, 233 images as our testing set. Moreover, we also verify the performance on Celeb A (Liu et al., 2015) which is a more challenging in the wild face images dataset. |
| Dataset Splits | Yes | In our experiments, we use the CASIA-Web Face dataset (Yi et al., 2014) with 494, 414 images as our training set and use the LFW dataset (Huang et al., 2008) with 13, 233 images as our testing set. Moreover, we also verify the performance on Celeb A (Liu et al., 2015) which is a more challenging in the wild face images dataset. |
| Hardware Specification | Yes | We implement our C-Face networks with the PyTorch framework and train them using four NVIDIA 1080Ti GPUs. |
| Software Dependencies | No | We implement our C-Face networks with the PyTorch framework and train them using four NVIDIA 1080Ti GPUs. (PyTorch is mentioned but without a version number) |
| Experiment Setup | Yes | We use Adam optimizer with a decayed learning rate of 0.99 and the batch size is 64 in every experiment. The learning rates of the GAN in stage 1 and stage 4 are 0.0002 and 1e-5 respectively. We train the GAN model for 10 epochs in stage 1 and 5 epochs in stage 4. The initial learning rate of stage 3 is 0.1 but it will be multiplied by 0.1 after each epoch and the CNNR will be fine-tuned for 10 epochs. We set α = 0.1 and β = 1.0 for (7) in each stage. For (6), we set γ1 = 0.05, γ2 = 0.1 in stage 1 and γ1 = γ2 = 0.05 in stage 4. |