Single image super-resolution

Description of the task

The goal of this workflow is to reconstruct high-resolution (HR) images from low-resolution (LR) ones. If there is a difference in the size of the LR and HR images, typically determined by a scale factor (x2, x4), this task is known as single-image super-resolution. If the size of the LR and HR images is the same, this task is usually referred to as image restoration.

An example of this task is displayed in the figure below, with a LR fluorescence microscopy image used as input (left) and its corresponding HR image (x2 scale factor).

../_images/LR_sr.png

LR fluorescence image from the
F-actin dataset by Qiao et al.
../_images/HR_sr.png

Corresponding HR image
at x2 resolution.

Notice that the LR image has been resized but its actual size is 502x502 pixels, whereas the size of its HR counterpart is 1004x1004.

Inputs and outputs

The super-resolution workflows in BiaPy expect a series of folders as input:

  • Training LR Images: A folder that contains the LR (single-channel or multi-channel) images that will be used to train the model.

    Expand to see how to configure

    In the current BiaPy GUI, this option is defined through the Wizard questions. Alternatively, you can edit the DATA.TRAIN.PATH in your YAML file before clicking Run Workflow and loading that YAML file.

  • Training HR Images: A folder that contains the HR (single- or multi-channel) images for training. Ensure their number match that of the training LR images.

    Expand to see how to configure

    In the current BiaPy GUI, this option is defined through the Wizard questions. Alternatively, you can edit the DATA.TRAIN.GT_PATH in your YAML file before clicking Run Workflow and loading that YAML file.

  • [Optional] Test LR Images: A folder that contains the images to evaluate the model's performance.
    Expand to see how to configure

    In the current BiaPy GUI, this option is defined through the Wizard questions. Alternatively, you can edit the DATA.TEST.PATH in your YAML file before clicking Run Workflow and loading that YAML file.

  • [Optional] Test HR Images: A folder that contains the HR images for testing. Again, ensure their count and sizes align with the test raw images.
    Expand to see how to configure

    In the current BiaPy GUI, this option is defined through the Wizard questions. Alternatively, you can edit the DATA.TEST.GT_PATH in your YAML file before clicking Run Workflow and loading that YAML file.

Upon successful execution, a directory will be generated with the segmentation results. Therefore, you will need to define:

  • Output Folder: A designated path to save the segmentation outcomes.

    Expand to see how to configure

    Under Run Workflow, click on the Browse button of Output folder to save the results:

    ../_images/GUI-run-workflow7.png
Graphical description of minimal inputs and outputs in BiaPy for super-resolution.

BiaPy input and output folders for super-resolution.

Data structure

To ensure the proper operation of the workflow, the directory tree should be something like this:

dataset/
β”œβ”€β”€ train
β”‚   β”œβ”€β”€ LR
β”‚   β”‚   β”œβ”€β”€ training_0001.tif
β”‚   β”‚   β”œβ”€β”€ training_0002.tif
β”‚   β”‚   β”œβ”€β”€ . . .
β”‚   β”‚   └── training_9999.tif
β”‚   └── HR
β”‚       β”œβ”€β”€ training_0001.tif
β”‚       β”œβ”€β”€ training_0002.tif
β”‚       β”œβ”€β”€ . . .
β”‚       └── training_9999.tif
└── test
    β”œβ”€β”€ LR
    β”‚   β”œβ”€β”€ testing_0001.tif
    β”‚   β”œβ”€β”€ testing_0002.tif
    β”‚   β”œβ”€β”€ . . .
    β”‚   └── testing_9999.tif
    └── HR
        β”œβ”€β”€ testing_0001.tif
        β”œβ”€β”€ testing_0002.tif
        β”œβ”€β”€ . . .
        └── testing_9999.tif

In this example, the LR training images are under dataset/train/LR/ and their corresponding HR images are under dataset/train/HR/, while the LR test images are under dataset/test/LR/ and their corresponding HR are under dataset/test/HR/. This is just an example, you can name your folders as you wish as long as you set the paths correctly later.

Note

Ensure that the LR and HR images are sorted in the same way. A common approach is to give the same name to each LR image and its corresponding HR image, or to fill with zeros the image number added to the filenames (as in the example).

Example datasets

Below is a list of publicly available datasets that are ready to be used in BiaPy for single image super-resolution:

Example dataset

Image dimensions

Link to data

F-actin dataset (ZeroCostDL4Mic)

2D

f_actin_sr_2d.zip

Confocal 2 STED - Nuclear Pore complex

3D

Nuclear_Pore_complez_3D.zip

Minimal configuration

Apart from the input and output folders, there are a few basic parameters that always need to be specified in order to run an super-resolution workflow in BiaPy. Depending on the parameter, they can be defined through the GUI Wizard, in the code-free notebooks, or by editing the YAML configuration file.

Experiment name

Also known as β€œmodel name” or β€œjob name”, this will be the name of the current experiment you want to run, so it can be differenciated from other past and future experiments.

Expand to see how to configure

Under Run Workflow, type the name you want for the job in the Job name field:

../_images/GUI-run-workflow7.png

Note

Use only my_model -style, not my-model (Use β€œ_” not β€œ-β€œ). Do not use spaces in the name. Avoid using the name of an existing experiment/model/job (saved in the same result folder) as it will be overwritten.

Data management

Validation Set

With the goal to monitor the training process, it is common to use a third dataset called the β€œValidation Set”. This is a subset of the whole dataset that is used to evaluate the model’s performance and optimize training parameters. This subset will not be directly used for training the model, and thus, when applying the model to these images, we can see if the model is learning the training set’s patterns too specifically or if it is generalizing properly.

Graphical description of data partitions in BiaPy

Graphical description of data partitions in BiaPy.

To define such set, there are two options:

  • Validation proportion/percentage: Select a proportion (or percentage) of your training dataset to be used to validate the network during the training. Usual values are 0.1 (10%) or 0.2 (20%), and the samples of that set will be selected at random.

    Expand to see how to configure

    In the current BiaPy GUI, this option is configured by editing the DATA.VAL.SPLIT_TRAIN in your YAML file before clicking Run Workflow and loading that YAML file.

  • Validation paths: Similar to the training and test sets, you can select two folders with the validation LR and HR images:

    • Validation LR Images: A folder that contains the unprocessed (single-channel or multi-channel) LR images that will be used to select the best model during training.

      Expand to see how to configure

      In the current BiaPy GUI, this option is configured by editing the DATA.VAL.PATH in your YAML file before clicking Run Workflow and loading that YAML file.

    • Validation HR Images: A folder that contains the instance (single-channel or multi-channel) HR images for validation. Ensure the number and ordering match those of the validation LR images.

      Expand to see how to configure

      In the current BiaPy GUI, this option is configured by editing the DATA.VAL.GT_PATH in your YAML file before clicking Run Workflow and loading that YAML file.

Test ground-truth

Do you have HR images for the test set? This is a key question so BiaPy knows if your test set will be used for evaluation in new data (unseen during training) or simply produce predictions on that new data. All workflows contain a parameter to specify this aspect.

Expand to see how to configure

In the current BiaPy GUI, this option is defined through the Wizard questions. Alternatively, you can edit the DATA.TEST.LOAD_GT in your YAML file before clicking Run Workflow and loading that YAML file.

Basic training parameters

At the core of each BiaPy workflow there is a deep learning model. Although we try to simplify the number of parameters to tune, these are the basic parameters that need to be defined for training a super-resolution workflow:

  • Number of input channels: The number of channels of your raw images (grayscale = 1, RGB = 3). Notice the dimensionality of your images (2D/3D) is set by default depending on the workflow template you select.

    Expand to see how to configure

    In the current BiaPy GUI, this option is configured by editing the DATA.PATCH_SIZE in your YAML file before clicking Run Workflow and loading that YAML file.

  • Scale factors: Factors by which the images will be super-resolved in X, Y and, if the images are 3D, in Z. If set to 1, the model will perform image restoration.

    Expand to see how to configure

    In the current BiaPy GUI, this option is configured by editing the PROBLEM.SUPER_RESOLUTION.UPSCALING in your YAML file before clicking Run Workflow and loading that YAML file.

  • Number of epochs: This number indicates how many rounds the network will be trained. On each round, the network usually sees the full training set. The value of this parameter depends on the size and complexity of each dataset. You can start with something like 100 epochs and tune it depending on how fast the loss (error) is reduced.

    Expand to see how to configure

    In the current BiaPy GUI, this option is configured by editing the TRAIN.EPOCHS in your YAML file before clicking Run Workflow and loading that YAML file.

  • Patience: This is a number that indicates how many epochs you want to wait without the model improving its results in the validation set to stop training. Again, this value depends on the data you’re working on, but you can start with something like 20.

    Expand to see how to configure

    In the current BiaPy GUI, this option is configured by editing the TRAIN.PATIENCE in your YAML file before clicking Run Workflow and loading that YAML file.

For improving performance, other advanced parameters can be optimized, for example, the model’s architecture. The architecture assigned as default is usually the RCAN, as it is effective in super-resolution tasks. This architecture allows a strong baseline, but further exploration could potentially lead to better results.

Note

Once the parameters are correctly assigned, the training phase can be executed. Note that to train large models effectively the use of a GPU (Graphics Processing Unit) is essential. This hardware accelerator performs parallel computations and has larger RAM memory compared to the CPUs, which enables faster training times.

How to run

BiaPy offers different options to run workflows depending on your degree of computer expertise. Select whichever is more approppriate for you:

In the BiaPy GUI, click on the Wizard, then follow the next instructions to select the single image super-resolution workflow:

After that, you will be able to edit the parameters of the workflow and run it.

Note

BiaPy’s GUI requires that all data and configuration files reside on the same machine where the GUI is being executed.

Tip

If you need additional help, watch BiaPy’s GUI walkthrough video.

Templates

In the templates/super-resolution folder of BiaPy, you will find a few YAML configuration templates for this workflow.

[Advanced] Special workflow configuration

Note

This section is recommended for experienced users only to improve the performance of their workflows. When in doubt, do not hesitate to check our FAQ & Troubleshooting or open a question in the image.sc discussion forum.

Advanced Parameters

Many of the parameters of our workflows are set by default to values that work commonly well. However, it may be needed to tune them to improve the results of the workflow. For instance, you may modify the following parameters:

  • Model architecture: Select the architecture of the deep neural network used as backbone of the pipeline. Options: EDSR, RCAN, WDSR, DFCAN, U-Net, Residual U-Net, Attention U-Net, SEUNet, MultiResUNet, ResUNet++, ResUNet SE and U-NeXt V1. Safe option: RCAN.

  • Batch size: This parameter defines the number of patches seen in each training step. Reducing or increasing the batch size may slow or speed up your training, respectively, and can influence network performance. Common values are 4, 8, 16, etc.

  • Patch size: Input the size of the patches use to train your model (length in pixels in X and Y). The value should be smaller or equal to the dimensions of the image. The default value is 256 in 2D, i.e. 256x256 pixels.

  • Optimizer: Select the optimizer used to train your model. Options: ADAM, ADAMW, Stochastic Gradient Descent (SGD). ADAM usually converges faster, while ADAMW provides a balance between fast convergence and better handling of weight decay regularization. SGD is known for better generalization. Default value: ADAMW.

  • Initial learning rate: Input the initial value to be used as learning rate. If you select ADAM as optimizer, this value should be around 10e-4.

  • Learning rate scheduler: Select to adjust the learning rate between epochs. The current options are β€œReduce on plateau”, β€œOne cycle”, β€œWarm-up cosine decay” or no scheduler.

  • Test time augmentation (TTA): Select to apply augmentation (flips and rotations) at test time. It usually provides more robust results but uses more time to produce each result. By default, no TTA is peformed.

Metrics

During the training and inference phases (if HR test images were provided, i.e. ground truth, and consequently, DATA.TEST.LOAD_GT is True) the performance of the model is measured using different metrics. Those metrics can be defined programmatically using the TRAIN.METRICS and TEST.METRICS variables of the YAML configuration file (a list of them is possible).

During training and test, the following metrics are available:

  • Peak signal-to-noise ratio (PSNR). Keyword: psnr.

  • Mean absolute error (MAE). Keyword: mae.

  • Mean squared error (MSE). Keyword: mse.

  • Structural similarity index measure (SSIM). Keyword: ssim.

Additionally, during test, if the images are 2D, the following metrics are also available:

  • FrΓ©chet inception distance (FID). Keyword: fid.

  • Inception score (IS). Keyword: is.

  • Learned perceptual image patch similarity (LPIPS). Keyword: lpips.

Results

The results are placed in results folder under --result_dir directory with the --name given. An example of this workflow is depicted below:

../_images/pred_sr.png

Predicted HR image.
../_images/HR_sr.png

Target HR image.

Here both images are of size 1004x1004.

Following the example, you should see that the directory /home/user/exp_results/my_2d_super_resolution has been created. If the same experiment is run 5 times, varying --run_id argument only, you should find the following directory tree:

Expand directory tree
my_2d_super_resolution/
β”œβ”€β”€ config_files
β”‚   └── 2d_super-resolution.yaml
β”œβ”€β”€ checkpoints
β”‚   └── my_2d_super-resolution_1-checkpoint-best.pth
└── results
    β”œβ”€β”€ my_2d_super_resolution_1
    β”œβ”€β”€ . . .
    └── my_2d_super_resolution_5
        β”œβ”€β”€ aug
        β”‚   └── .tif files
        β”œβ”€β”€ charts
        β”‚   β”œβ”€β”€ my_2d_super_resolution_1_*.png
        β”‚   └── my_2d_super_resolution_1_loss.png
        β”œβ”€β”€ per_image
        β”‚   β”œβ”€β”€ .tif files
        β”‚   └── .zarr files (or.h5)
        β”œβ”€β”€ train_logs
        └── tensorboard

  • config_files: directory where the .yaml filed used in the experiment is stored.

    • 2d_super-resolution.yaml: YAML configuration file used (it will be overwrited every time the code is run)

  • checkpoints, optional: directory where model’s weights are stored. Only created when TRAIN.ENABLE is True and the model is trained for at least one epoch.

    • my_2d_super-resolution_1-checkpoint-best.pth, optional: checkpoint file (best in validation) where the model’s weights are stored among other information. Only created when the model is trained for at least one epoch.

    • normalization_mean_value.npy, optional: normalization mean value. Is saved to not calculate it everytime and to use it in inference. Only created if DATA.NORMALIZATION.TYPE is custom.

    • normalization_std_value.npy, optional: normalization std value. Is saved to not calculate it everytime and to use it in inference. Only created if DATA.NORMALIZATION.TYPE is custom.

  • results: directory where all the generated checks and results will be stored. There, one folder per each run are going to be placed.

    • my_2d_super_resolution_1: run 1 experiment folder. Can contain:

      • aug, optional: image augmentation samples. Only created if AUGMENTOR.AUG_SAMPLES is True.

      • charts, optional: only created when TRAIN.ENABLE is True and epochs trained are more or equal LOG.CHART_CREATION_FREQ. Can contain:

        • my_2d_super_resolution_1_*.png: Plot of each metric used during training.

        • my_2d_super_resolution_1_loss.png: Loss over epochs plot.

      • per_image:

        • .tif files, optional: reconstructed images from patches. Created when TEST.BY_CHUNKS.ENABLE is False or when TEST.BY_CHUNKS.ENABLE is True but TEST.BY_CHUNKS.SAVE_OUT_TIF is True.

        • .zarr files (or.h5), optional: reconstructed images from patches. Created when TEST.BY_CHUNKS.ENABLE is True.

      • train_logs: each row represents a summary of each epoch stats. Only avaialable if training was done.

      • tensorboard: Tensorboard logs.

      • test_results_metrics.csv: a CSV file containing all the evaluation metrics obtained on each file of the test set if ground truth was provided.

Note

Here, for visualization purposes, only my_2d_super_resolution_1 has been described but my_2d_super_resolution_2, my_2d_super_resolution_3, my_2d_super_resolution_4 and my_2d_super_resolution_5 will follow the same structure.