""" Jimut Bahan Pal 31-May-2021 A Script to Augment the Smear Slides Cropped Dataset. """ import os import cv2 import json import glob import math import random import shutil import argparse import numpy as np from tqdm import tqdm from lxml import etree from tqdm import tqdm from PIL import Image, ImageDraw import matplotlib.pyplot as plt from scipy.interpolate import UnivariateSpline from skimage.util import random_noise print("LOADING BG TEXTURES...") NORMAL_BG_NAMES = glob.glob('normal_bg/*.jpg') STAIN_BG_NAMES = glob.glob('stain_bg/*.jpg') def _create_LUT_8UC1(x, y): spl = UnivariateSpline(x, y) return spl(range(256)) def get_random_vector(): # to generate a random one hot vector of size # positions are given by vectors # 0 - apply_normal_background_image(img) or apply_stain_background_image(img) # 1 - zoom_out_image(img, zoom_percentage) # 2 - flip_image(img, opt) # 3 - apply_color_transformation_image(img, val1, val2, val3) # 4 - gaussian_blur_repeatitive_image(img, kernel_size,n_iter) # 5 - rotate_image(img, angle) # 6 - change_contrast_image(img, alpha) # 7 - change_brightness_image(img, beta) # 8 - apply_noise_gaussian_image(img) # 9 - apply_noise_salt_and_pepper(img, amount) random_vector = [] for i in range(11): random_vector.append(random.randint(0,1)) print(random_vector) print(len(random_vector)) return random_vector # some of the functions to do image augmentation for augmenting the entire # dataset for performing domain adaptation # This functions should be added before any step def apply_normal_background_image(img): # print(img[0][0][0]) background_image = cv2.imread(NORMAL_BG_NAMES[random.randint(0,(len(NORMAL_BG_NAMES)-1))]) background_image = cv2.cvtColor(background_image, cv2.COLOR_BGR2RGB) im_h, im_w, n_c = img.shape background_image = cv2.resize(background_image,(im_w,im_h)) # print(background_image.shape) # print(img.shape) # median blur to remove noises img_blurred = cv2.medianBlur(img, 111) background_image[img_blurred > 0] = 0 background_added_image = background_image + img return background_added_image # This functions should be added before any step def apply_stain_background_image(img): # apply background with possible stain background_image = cv2.imread(STAIN_BG_NAMES[random.randint(0,(len(STAIN_BG_NAMES)-1))]) background_image = cv2.cvtColor(background_image, cv2.COLOR_BGR2RGB) im_h, im_w, n_c = img.shape background_image = cv2.resize(background_image,(im_w,im_h)) # print(background_image.shape) # print(img.shape) # median blur to remove noises img_blurred = cv2.medianBlur(img, 111) background_image[img_blurred > 0] = 0 background_added_image = background_image + img return background_added_image def zoom_out_image(img, zoom_percentage): # zooms out an image by certain percentage # please keep this between 0.5 - 0.9 im_h, im_w, im_c = img.shape resized_height, resized_width = int(zoom_percentage*im_h), int(zoom_percentage*im_w) # print(resized_height,",",resized_width) img = cv2.resize(img, (resized_height,resized_width)) # make the image to original size delta_w = im_h - resized_height delta_h = im_w - resized_width top, bottom = delta_h//2, delta_h-(delta_h//2) left, right = delta_w//2, delta_w-(delta_w//2) color = [0, 0, 0] new_im = cv2.copyMakeBorder(img, top, bottom, left, right, cv2.BORDER_CONSTANT, value=color) return new_im def flip_image(img, opt): # filp an image = horizontally or vertically or both # 0 for horizontal flip, 1 for vertical flip, -1 for both return cv2.flip(img, opt) def apply_color_transformation_image(img, val1, val2, val3): # apply colour transformation to image, generally bluish by certain variant # or transform an image to certain colour, specially with a tinge of yellow, and blue # this can also be considered as a warm filter, and a cold filter changed according to # value decr_ch_lut = _create_LUT_8UC1(val1, val2) incr_ch_lut = _create_LUT_8UC1(val1, val3) c_r, c_g, c_b = cv2.split(img) c_r = cv2.LUT(c_r, decr_ch_lut).astype(np.uint8) c_b = cv2.LUT(c_b, incr_ch_lut).astype(np.uint8) img_rgb = cv2.merge((c_r, c_g, c_b)) # decrease color saturation c_h, c_s, c_v = cv2.split(cv2.cvtColor(img_rgb, cv2.COLOR_RGB2HSV)) c_s = cv2.LUT(c_s, decr_ch_lut).astype(np.uint8) return cv2.cvtColor(cv2.merge((c_h, c_s, c_v)), cv2.COLOR_HSV2RGB) def gaussian_blur_repeatitive_image(img, kernel_size,n_iter): # apply gaussian blur with different kernel size and # repeat it different times in an image for iter in range(n_iter): img = cv2.GaussianBlur(img, kernel_size, cv2.BORDER_DEFAULT) return img def rotate_image(img, angle): # rotate an image by certain angle row,col,channel = img.shape center=tuple(np.array([row,col])/2) rot_mat = cv2.getRotationMatrix2D(center,angle,1.0) new_image = cv2.warpAffine(img, rot_mat, (col,row)) return new_image # def zoom_in_image(img, zoom_percentage): # # zoom in an image by certain percentage, doesn't make # # any sense here... # pass def change_contrast_image(img, alpha): # change contrast of an image # alpha = 1.5 # Contrast control (1.0-3.0) beta = 0 # Brightness control (0-100) adjusted = cv2.convertScaleAbs(img, alpha=alpha, beta=beta) return adjusted def change_brightness_image(img, beta): # Change brightness of an image alpha = 1.5 # Contrast control (1.0-3.0) # beta = 0 # Brightness control (0-100) adjusted = cv2.convertScaleAbs(img, alpha=alpha, beta=beta) return adjusted def apply_noise_gaussian_image(img): # apply salt and pepper noise in an image noise_img = random_noise(img, mode='gaussian', seed=None, clip=True) # The above function returns a floating-point image # on the range [0, 1], thus we changed it to 'uint8' # and from [0,255] noise_img = np.array(255*noise_img, dtype = 'uint8') # print(noise_img.max(), noise_img.min()) return noise_img def apply_noise_salt_and_pepper(img, amount): # apply salt and pepper noise to the image noise_img = random_noise(img, mode='s&p',amount=0.3) # The above function returns a floating-point image # on the range [0, 1], thus we changed it to 'uint8' # and from [0,255] noise_img = np.array(255*noise_img, dtype = 'uint8') return noise_img def show_img(img): plt.imshow(img) plt.show() folders = glob.glob('classification_data/*') print(folders) for folder in folders: image_files = glob.glob('{}/*'.format(folder)) for image_name in image_files: if 'yml' not in image_name: image = cv2.imread(image_name,cv2.IMREAD_COLOR) for iter_ in tqdm(range(100)): vector_ = get_random_vector() dummy_count = 0 for choice in vector_: # make the choices according to the vector if choice == 0 and dummy_count == 0: # apply normal background to the image image_out = apply_normal_background_image(image) if choice == 1 and dummy_count == 0: # apply stain background to the image image_out = apply_stain_background_image(image) if choice == 1 and dummy_count == 1: # apply zoom out on an image # generates random value between 0.49 to 1.0 # probably will be better if this is aliased with other function gen_random = (random.randint(7,10)/10)*(random.randint(7,10)/10) image_out = zoom_out_image(image_out, gen_random) if choice == 1 and dummy_count == 2: image_out = flip_image(image_out, opt=random.randint(-1,1)) if choice == 1 and dummy_count == 3: get_ch = random.randint(0,5) # applying cooler transformations if get_ch == 0: image_out = apply_color_transformation_image(image_out, [0, 64, 128, 192, 256], [0, 30, 80, 120, 192], [0, 40, 95, 142.5, 208]) if get_ch == 1: image_out = apply_color_transformation_image(image_out, [0, 64, 128, 192, 256], [0, 30, 80, 120, 192], [0, 70, 140, 210, 256]) if get_ch == 2: image_out = apply_color_transformation_image(image_out, [0, 64, 128, 192, 256], [0, 70, 100, 140, 192], [0, 110, 180, 210, 256]) # apply warmer transofrmations if get_ch == 3: image_out = apply_color_transformation_image(image_out, [0, 64, 128, 192, 256], [0, 70, 100, 140, 192], [0, 110, 180, 210, 256]) if get_ch == 4: image_out = apply_color_transformation_image(image_out, [0, 64, 128, 192, 256], [0, 70, 140, 210, 256], [0, 30, 80, 120, 192]) if get_ch == 5: image_out = apply_color_transformation_image(image_out, [0, 64, 128, 192, 256], [0, 110, 180, 210, 256], [0, 70, 100, 140, 192]) if choice == 1 and dummy_count == 4: # apply gaussian blur on images till limit get_ch = random.randint(0,2) if get_ch == 0: image_out = gaussian_blur_repeatitive_image(image_out, (5,5), 15) if get_ch == 1: image_out = gaussian_blur_repeatitive_image(image_out, (11,11), 15) if get_ch == 2: image_out = gaussian_blur_repeatitive_image(image_out, (17,17), 15) if choice == 1 and dummy_count == 5: # rotate an image by certain angle get_ch = random.randint(0,3) if get_ch == 0: image_out = rotate_image(image_out, random.randint(0,90)) if get_ch == 1: image_out = rotate_image(image_out, random.randint(90,180)) if get_ch == 2: image_out = rotate_image(image_out, random.randint(180,270)) if get_ch == 3: image_out = rotate_image(image_out, random.randint(270,360)) if choice == 1 and dummy_count == 6: # change the contrast of the image # change alpha from (1.0-3.0) by a random number get_random = (random.randint(100,173)/100) * (random.randint(100,173)/100) image_out = change_contrast_image(image_out, get_random) if choice == 1 and dummy_count == 7: # Change the brightness of the image image_out = change_brightness_image(image_out, random.randint(1,75)) pass if choice == 1 and dummy_count == 8: # apply Gaussian Noise image_out = apply_noise_gaussian_image(image_out) if choice == 1 and dummy_count == 9: # apply salt and pepper noise amount = 1 image_out = apply_noise_salt_and_pepper(image_out, amount) dummy_count += 1 # show_img(image_out) save_name = image_name.split('.')[0]+"_"+str(iter_)+".jpg" cv2.imwrite(save_name,image_out)