drln-bam / README.md
Eugene Siow
Add x2, x3 model. c5d6eff
1 ---
2 license: apache-2.0
3 tags:
4 - super-image
5 - image-super-resolution
6 datasets:
7 - eugenesiow/Div2k
8 - eugenesiow/Set5
9 - eugenesiow/Set14
10 - eugenesiow/BSD100
11 - eugenesiow/Urban100
12 metrics:
13 - pnsr
14 - ssim
15 ---
16 # Densely Residual Laplacian Super-Resolution (DRLN)
17 DRLN model pre-trained on DIV2K (800 images training, augmented to 4000 images, 100 images validation) for 2x, 3x and 4x image super resolution. It was introduced in the paper [Densely Residual Laplacian Super-resolution](https://arxiv.org/abs/1906.12021) by Anwar et al. (2020) and first released in [this repository](https://github.com/saeed-anwar/DRLN).
18
19 The goal of image super resolution is to restore a high resolution (HR) image from a single low resolution (LR) image. The image below shows the ground truth (HR), the bicubic upscaling and model upscaling.
20
21 ![Comparing Bicubic upscaling against the models x4 upscaling on Set5 Image 4](images/drln_4_4_compare.png "Comparing Bicubic upscaling against the models x4 upscaling on Set5 Image 4")
22 ## Model description
23 Super-Resolution convolutional neural networks have recently demonstrated high-quality restoration for single images. However, existing algorithms often require very deep architectures and long training times. Furthermore, current convolutional neural networks for super-resolution are unable to exploit features at multiple scales and weigh them equally, limiting their learning capability. In this exposition, we present a compact and accurate super-resolution algorithm namely, Densely Residual Laplacian Network (DRLN). The proposed network employs cascading residual on the residual structure to allow the flow of low-frequency information to focus on learning high and mid-level features. In addition, deep supervision is achieved via the densely concatenated residual blocks settings, which also helps in learning from high-level complex features. Moreover, we propose Laplacian attention to model the crucial features to learn the inter and intra-level dependencies between the feature maps. Furthermore, comprehensive quantitative and qualitative evaluations on low-resolution, noisy low-resolution, and real historical image benchmark datasets illustrate that our DRLN algorithm performs favorably against the state-of-the-art methods visually and accurately.
24
25 This model also applies the balanced attention (BAM) method invented by [Wang et al. (2021)](https://arxiv.org/abs/2104.07566) to further improve the results.
26 ## Intended uses & limitations
27 You can use the pre-trained models for upscaling your images 2x, 3x and 4x. You can also use the trainer to train a model on your own dataset.
28 ### How to use
29 The model can be used with the [super_image](https://github.com/eugenesiow/super-image) library:
30 ```bash
31 pip install super-image
32 ```
33 Here is how to use a pre-trained model to upscale your image:
34 ```python
35 from super_image import DrlnModel, ImageLoader
36 from PIL import Image
37 import requests
38
39 url = 'https://paperswithcode.com/media/datasets/Set5-0000002728-07a9793f_zA3bDjj.jpg'
40 image = Image.open(requests.get(url, stream=True).raw)
41
42 model = DrlnModel.from_pretrained('eugenesiow/drln-bam', scale=2) # scale 2, 3 and 4 models available
43 inputs = ImageLoader.load_image(image)
44 preds = model(inputs)
45
46 ImageLoader.save_image(preds, './scaled_2x.png') # save the output 2x scaled image to `./scaled_2x.png`
47 ImageLoader.save_compare(inputs, preds, './scaled_2x_compare.png') # save an output comparing the super-image with a bicubic scaling
48 ```
49 [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/eugenesiow/super-image-notebooks/blob/master/notebooks/Upscale_Images_with_Pretrained_super_image_Models.ipynb "Open in Colab")
50 ## Training data
51 The models for 2x, 3x and 4x image super resolution were pretrained on [DIV2K](https://huggingface.co/datasets/eugenesiow/Div2k), a dataset of 800 high-quality (2K resolution) images for training, augmented to 4000 images and uses a dev set of 100 validation images (images numbered 801 to 900).
52 ## Training procedure
53 ### Preprocessing
54 We follow the pre-processing and training method of [Wang et al.](https://arxiv.org/abs/2104.07566).
55 Low Resolution (LR) images are created by using bicubic interpolation as the resizing method to reduce the size of the High Resolution (HR) images by x2, x3 and x4 times.
56 During training, RGB patches with size of 64×64 from the LR input are used together with their corresponding HR patches.
57 Data augmentation is applied to the training set in the pre-processing stage where five images are created from the four corners and center of the original image.
58
59 We need the huggingface [datasets](https://huggingface.co/datasets?filter=task_ids:other-other-image-super-resolution) library to download the data:
60 ```bash
61 pip install datasets
62 ```
63 The following code gets the data and preprocesses/augments the data.
64
65 ```python
66 from datasets import load_dataset
67 from super_image.data import EvalDataset, TrainDataset, augment_five_crop
68
69 augmented_dataset = load_dataset('eugenesiow/Div2k', 'bicubic_x4', split='train')\
70 .map(augment_five_crop, batched=True, desc="Augmenting Dataset") # download and augment the data with the five_crop method
71 train_dataset = TrainDataset(augmented_dataset) # prepare the train dataset for loading PyTorch DataLoader
72 eval_dataset = EvalDataset(load_dataset('eugenesiow/Div2k', 'bicubic_x4', split='validation')) # prepare the eval dataset for the PyTorch DataLoader
73 ```
74 ### Pretraining
75 The model was trained on GPU. The training code is provided below:
76 ```python
77 from super_image import Trainer, TrainingArguments, DrlnModel, DrlnConfig
78
79 training_args = TrainingArguments(
80 output_dir='./results', # output directory
81 num_train_epochs=1000, # total number of training epochs
82 )
83
84 config = DrlnConfig(
85 scale=4, # train a model to upscale 4x
86 bam=True, # apply balanced attention to the network
87 )
88 model = DrlnModel(config)
89
90 trainer = Trainer(
91 model=model, # the instantiated model to be trained
92 args=training_args, # training arguments, defined above
93 train_dataset=train_dataset, # training dataset
94 eval_dataset=eval_dataset # evaluation dataset
95 )
96
97 trainer.train()
98 ```
99
100 [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/eugenesiow/super-image-notebooks/blob/master/notebooks/Train_super_image_Models.ipynb "Open in Colab")
101 ## Evaluation results
102 The evaluation metrics include [PSNR](https://en.wikipedia.org/wiki/Peak_signal-to-noise_ratio#Quality_estimation_with_PSNR) and [SSIM](https://en.wikipedia.org/wiki/Structural_similarity#Algorithm).
103
104 Evaluation datasets include:
105 - Set5 - [Bevilacqua et al. (2012)](https://huggingface.co/datasets/eugenesiow/Set5)
106 - Set14 - [Zeyde et al. (2010)](https://huggingface.co/datasets/eugenesiow/Set14)
107 - BSD100 - [Martin et al. (2001)](https://huggingface.co/datasets/eugenesiow/BSD100)
108 - Urban100 - [Huang et al. (2015)](https://huggingface.co/datasets/eugenesiow/Urban100)
109
110 The results columns below are represented below as `PSNR/SSIM`. They are compared against a Bicubic baseline.
111
112 |Dataset |Scale |Bicubic |drln-bam |
113 |--- |--- |--- |--- |
114 |Set5 |2x |33.64/0.9292 |**38.23/0.9614** |
115 |Set5 |3x |30.39/0.8678 |**35.3/0.9422** |
116 |Set5 |4x |28.42/0.8101 |**32.49/0.8986** |
117 |Set14 |2x |30.22/0.8683 |**33.95/0.9206** |
118 |Set14 |3x |27.53/0.7737 |**31.27/0.8624** |
119 |Set14 |4x |25.99/0.7023 |**28.94/0.7899** |
120 |BSD100 |2x |29.55/0.8425 |**33.95/0.9269** |
121 |BSD100 |3x |27.20/0.7382 |**29.78/0.8224** |
122 |BSD100 |4x |25.96/0.6672 |**28.63/0.7686** |
123 |Urban100 |2x |26.66/0.8408 |**32.81/0.9339** |
124 |Urban100 |3x | |**29.82/0.8828** |
125 |Urban100 |4x |23.14/0.6573 |**26.53/0.7991** |
126
127 ![Comparing Bicubic upscaling against the models x4 upscaling on Set5 Image 2](images/drln_2_4_compare.png "Comparing Bicubic upscaling against the models x4 upscaling on Set5 Image 2")
128
129 You can find a notebook to easily run evaluation on pretrained models below:
130
131 [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/eugenesiow/super-image-notebooks/blob/master/notebooks/Evaluate_Pretrained_super_image_Models.ipynb "Open in Colab")
132
133 ## BibTeX entry and citation info
134 ```bibtex
135 @misc{wang2021bam,
136 title={BAM: A Lightweight and Efficient Balanced Attention Mechanism for Single Image Super Resolution},
137 author={Fanyi Wang and Haotian Hu and Cheng Shen},
138 year={2021},
139 eprint={2104.07566},
140 archivePrefix={arXiv},
141 primaryClass={eess.IV}
142 }
143 ```
144
145 ```bibtex
146 @misc{anwar2019densely,
147 title={Densely Residual Laplacian Super-Resolution},
148 author={Saeed Anwar and Nick Barnes},
149 year={2019},
150 eprint={1906.12021},
151 archivePrefix={arXiv},
152 primaryClass={eess.IV}
153 }
154 ```