AI Tools for the DAITA Platform

Preprocessing and Augmentation

Preprocessing Methods

List of preprocessing methods:

• RotateExif: Rotate image according to Exif information.
• Grayscale: Grayscale an image.
• NormalizeBrightness: Adjust brightness of image to match with a reference image if the difference > 75%.
• NormalizeHue: Adjust hue of image to match with a reference image if the difference > 75%.
• NormalizeSaturation: Adjust saturation of image to match with a reference image if the difference > 75%.
• NormalizeSharpness: Adjust sharpness of image to match with a reference image if the difference > 75%.
• NormalizeContrast: Adjust contrast of image to match with a reference image if the difference > 75%.
• EqualizeHistogram: Equalize the histogram of an image if the contrast is low.
• IncreaseResolution: Increase resolution of a tensor image given a reference image.

Augmentation Methods

List of augmentation methods:

• random_rotate: Rotate an image with random angle. Degree range: -20 to 20.
• random_scale: Zoom into a random region in an image. Scaling factor range: 0.5 to 2.
• random_translate: Shift image randomly in vertical direction or horizontal direction. Shifting percentage: -img_width * a < dx < img_width * a or -img_height * b < dx < img_width * b, a = 0.2 and b = 0.2.
• random_horizontal_flip: Flip image horizontally.
• random_vertical_flip: Flip image vertically.
• random_crop: Crop a random region in an image. Crop size: 512 x 512.
• random_tile: Tile an image and extract a random patch. Patch size: 512 x 512.
• random_erase: Choose a random region in image and fill it with black. Region's scale range: 0.02 to 0.3. Region's ratio range: 0.3 to 3.3.
• random_gaussian_noise: Add gaussian noise to image; mean = 0, standard deviation = 0.1.
• random_gaussian_blur: Blur an image with random kernel size and sigma. Kernel size range: 3 to 27. Sigma range: 1 to 10.
• random_sharpness: Enhance sharpness of image randomly. Sharpness strength: 0.5.
• random_brightness: Changing brightness of image randomly. Brightness factor range: 0.75 to 1.5.
• random_hue: Change hue of image randomly. Hue factor range: -0.5 to 0.5.
• random_saturation: Change saturation of image randomly. Saturation factor range: 0.5 to 1.5.
• random_contrast: Change contrast of image randomly. Contrast factor range: 0.5 to 1.5.
• random_solarize: Solarize an image randomly; thresholds = 0.1, additions = 0.1.
• random_posterize: Posterize an image randomly; bits = 3.
• super_resolution: Change resolution of an image; scale = 2.

Installation

We recommend using Python versions 3.8 or 3.9. First, create a virtual environment:

virtualenv venv

The virtualenv package is required to create virtual environments; you can install it with pip: pip install virtualenv.

Now, activate the virtual environment:

• Mac OS / Linux: source venv/bin/activate
• Windows: venv/Scripts/activate

Eventually, install all dependencies:

pip install -r AI/requirements.txt --find-links=https://download.pytorch.org/whl/torch_stable.html

Deployment via Rest API on Endpoint 0.0.0.0:8000/ai

OMP_NUM_THREADS=12 ray start --head; python AI/deploy.py

How to Use

Preprocessing Methods

import requests

# Send HTTP POST request
response = requests.post(
    "http://0.0.0.0:8000/ai",
    json={
        "images_paths": [
            "AI/data/sample/images/20180810150607_camera_frontcenter_000000083.png",
            "AI/data/sample/images/20181016125231_camera_frontcenter_000183553.png",
            "AI/data/sample/images/20181108084007_camera_frontcenter_000029570.png",
            "AI/data/sample/images/20181108123750_camera_frontcenter_000004559.png",
            "AI/data/sample/images/20190401145936_camera_frontcenter_000017970.png"
        ],
        "output_folder": "AI/data/sample/output/",
        "type": "preprocessing",
        "codes": [],
    }
)
print(response.json())

Augmentation Methods

import requests

# Send HTTP POST request
response = requests.post(
    "http://0.0.0.0:8000/ai",
    json={
        "images_paths": [
            "AI/data/sample/images/20180810150607_camera_frontcenter_000000083.png",
            "AI/data/sample/images/20181016125231_camera_frontcenter_000183553.png",
            "AI/data/sample/images/20181108084007_camera_frontcenter_000029570.png",
            "AI/data/sample/images/20181108123750_camera_frontcenter_000004559.png",
            "AI/data/sample/images/20190401145936_camera_frontcenter_000017970.png"
        ],
        "output_folder": "AI/data/sample/output/",
        "type": "augmentation",
        "num_augments_per_image": 2,
        "codes": [],
    }
)
print(response.json())

Libraries

• Preprocessing on CPU: Scikit-image
• Augmentation on GPU and CPU: Kornia
• Deploy on REST API: Ray

A Note on Releases

The repository should be setup in the future in such a way that changes to the branch release-staging and release-production will trigger a rebuild of the Staging and the Production application, respectively.

The merge flow for changes should be as follows:

<feature_branch> -> <develop> -> <main> -> <release-staging> -> <release-production>