img_ai_prep

Image-based machine learning models often require square images of uniform size. Existing code to process images for ingestion into ML models often use naively cropped images to save on computation costs, but what about in data-constrained environments such as fine-tuning? This library helps you to more intelligently crop photos for ingestion to your ML models, to improve data efficiency. It supports two major workflows that will be useful to produce higher quality images:

resize_center_crop

1. Resize the image with the shortest side matching the desired end size
2. Crop only the longer side to achieve final square image, preserving maximum content towards the center

face_center_crop

1. Resize the image with the shortest side matching the desired end size
2. Crop towards detected faces in the image, preserving maximum human content

In addition to these enhanced cropping modes, the code also supports cropping towards faces without resizing, center cropping without resizing, and exposes the underlying crop and resize functions for use in your own workflows.

There is also support for easily passing custom models for face detection, and passing custom parameters to the models. Passing the model by argument also improves performance if you are working in a loop, as it does not need to read the model from disk every time if passed by argument.

Usage

Original Image	noresize_center_crop	resize_center_crop	face_center_crop	noresize_face_center

Simply install with pip and get to processing your images!

pip install img_ai_prep

from PIL import Image

from img_ai_prep import img_ai_prep

im = Image.open("input.jpg")
newimg = img_ai_prep(im,
                     final_size=1024,
                     crop_mode="resize_center_crop")
print(newimg.size)
newimg.save("output.jpg")

Advanced example script passing a custom model and parameters:

import cv2
from PIL import Image

from img_ai_prep import img_ai_prep

if __name__ == "__main__":
    im = Image.open("my_img.jpg")

    model = cv2.CascadeClassifier(cv2.data.haarcascades +
                                  "haarcascade_frontalface_alt.xml")
    newimg = img_ai_prep(im,
                         final_size=1024,
                         crop_mode="face_center_crop",
                         model=model,
                         min_neighbors=50,
                         min_size=(100,100))
    print(newimg.size)
    newimg.save("my_output.jpg")