This project requires data to be stored in a TensorFlow Datasets container.
Performing EDA on the original dataset revealed a wide range of aspect-ratios and raw image sizes. Since the ResNet50 model that we'll be using as the backbone network for this project has a fixed input dimension of 224 x 224 x 3, in order to minimize distortion during training and inference, we cropped images that were either extremely tall, or extremely wide. We also normalized images to a height of 800 pixels.
Here are the steps to generate the TFDS Dataset:
- CheckFiles-XRay.ipynb : Perform EDA on original images to determine thresholds for cropping and target image size
- Resize_XRay.ipynb : Perform Cropping and Resizing
- CheckFiles-XRay-Cropped.ipynb : Check results of Cropping and Resizing
- Generate_TFDS_Dataset.ipynb : Generate TFDS Dataset