| """ |
| image_augmentation.py |
| |
| This script performs data augmentation on images using various transformations such as color jitter, |
| Gaussian blur, sharpness adjustment, horizontal flip, and vertical flip. |
| It generates augmented images from an existing dataset and saves them to new folders. |
| |
| """ |
|
|
| from PIL import Image |
| import os |
| from torchvision.utils import save_image |
| import torchvision.transforms as tvts |
| import random |
|
|
| |
| transforms=[tvts.ColorJitter(brightness=.5, hue=.3), |
| tvts.GaussianBlur(kernel_size=(5, 9), sigma=(0.1, 5.)), |
| tvts.RandomAdjustSharpness(sharpness_factor=2), |
| tvts.RandomHorizontalFlip(p=0.5), |
| tvts.RandomVerticalFlip(p=0.5)] |
|
|
| def select_random_imgs(img_list, count): |
| """ |
| Selects a random subset of images from a list. |
| |
| Parameters: |
| img_list (list): List of image filenames. |
| count (int): Number of images to select. |
| |
| Returns: |
| list: Randomly selected image filenames. |
| """ |
| return random.sample(img_list, count) |
|
|
| def create_foldername(fname): |
| """ |
| Creates a new folder name for augmented images. |
| |
| Parameters: |
| fname (str): Original folder name. |
| |
| Returns: |
| str: Augmented folder name. |
| """ |
| names = fname.split("_") |
| names.insert(2, "augment") |
| return "_".join(names) |
|
|
| def save_original_imgs(source_fname, dest_fname, img_files): |
| """ |
| Saves original images to the destination folder. |
| |
| Parameters: |
| source_fname (str): Source folder name. |
| dest_fname (str): Destination folder name. |
| img_files (list): List of image filenames. |
| """ |
| for img in img_files: |
| img_name = img.split(".")[0] |
| img = open_PILimg(path=os.path.join(source_fname, img)) |
| save_image(tvts.ToTensor()(img), os.path.join(dest_fname, "{}.jpg".format(img_name))) |
|
|
| def open_PILimg(path): |
| """ |
| Opens an image file using PIL. |
| |
| Parameters: |
| path (str): File path to the image. |
| |
| Returns: |
| PIL.Image.Image: PIL image object. |
| """ |
| return Image.open(path).convert("RGB") |
| |
| def augment_folder(root_fname, fname, augment_count=25): |
| """ |
| Augments images in a folder and saves the augmented images to a new folder. |
| |
| Parameters: |
| root_fname (str): Root folder containing the source images. |
| fname (str): Name of the folder containing the source images. |
| augment_count (int): Number of augmented images to generate for each original image. |
| """ |
| onlyfiles = os.listdir(os.path.join(root_fname, fname)) |
| print(f'Befor augmentation, number of images: {len(onlyfiles)}') |
| random_imgs = select_random_imgs(onlyfiles, len(onlyfiles)) |
| augment_fname = create_foldername(root_fname) |
| if not os.path.exists(augment_fname): |
| os.mkdir(augment_fname) |
| |
| save_fname = os.path.join(augment_fname, fname) |
|
|
| if not os.path.exists(save_fname): |
| os.mkdir(save_fname) |
|
|
| save_original_imgs(source_fname=os.path.join(root_fname, fname), |
| dest_fname=os.path.join(augment_fname, fname), |
| img_files=onlyfiles) |
|
|
| for img in random_imgs: |
| img_name = img.split(".")[0] |
| img = open_PILimg(path=os.path.join(root_fname, fname, img)) |
| |
| for transform in transforms: |
| transname = transform.__class__.__name__ |
| transformed_imgs = [transform(img) for _ in range(augment_count)] |
| for i, trans_img in enumerate(transformed_imgs): |
| save_image(tvts.ToTensor()(trans_img), os.path.join(save_fname, "trans_img_{}_{}_{}.jpg".format(img_name, transname, i))) |
|
|
| print(f"After augmentation, number of images: {len(os.listdir(save_fname))}") |
| |
|
|
| if __name__ == "__main__": |
| |
| root_folder = [["GNSS/Train"]] |
| for root in root_folder: |
| folders = [f for f in os.listdir(root) if os.path.isdir(os.path.join(root, f))] |
| for folder in folders: |
| augment_folder(root, folder) |
|
|
|
|
|
|
