Parcourir la source

Merge branch 'release/v0.4.2'

Jérôme BUISINE il y a 4 ans
Parent
commit
e1dd8b4336
2 fichiers modifiés avec 217 ajouts et 0 suppressions
  1. 1 0
      custom_config.py
  2. 216 0
      generate/generate_data_augmentation.py

Fichier diff supprimé car celui-ci est trop grand
+ 1 - 0
custom_config.py


+ 216 - 0
generate/generate_data_augmentation.py

@@ -0,0 +1,216 @@
+# main imports
+import sys, os, argparse
+import numpy as np
+import time
+import random
+import math
+
+# image processing imports
+from PIL import Image
+
+from ipfml.processing import transform, segmentation
+from ipfml import utils
+
+# modules imports
+sys.path.insert(0, '') # trick to enable import of main folder module
+
+import custom_config as cfg
+from modules.utils import data as dt
+
+
+# getting configuration information
+zone_folder             = cfg.zone_folder
+min_max_filename        = cfg.min_max_filename_extension
+
+# define all scenes values
+scenes_list             = cfg.scenes_names
+scenes_indexes          = cfg.scenes_indices
+choices                 = cfg.normalization_choices
+path                    = cfg.dataset_path
+zones                   = cfg.zones_indices
+seuil_expe_filename     = cfg.seuil_expe_filename
+
+features_choices        = cfg.features_choices_labels
+output_data_folder      = cfg.output_data_folder
+
+
+image_scene_size        = (800, 800)
+image_zone_size         = (200, 200)
+possible_point_zone     = tuple(np.asarray(image_scene_size) - np.array(image_zone_size))
+data_augmented_filename = cfg.data_augmented_filename
+
+def main():
+    
+    parser = argparse.ArgumentParser(description="Compute and prepare data augmentation of scenes")
+
+    parser.add_argument('--output', type=str, help="output folder expected", required=True)
+    parser.add_argument('--number', type=int, help="number of images for each sample of scene", required=True)
+    parser.add_argument('--rotation', type=bool, help="", required=True, default=False)
+
+    args = parser.parse_args()
+
+    p_output   = args.output
+    p_number   = args.number
+    p_rotation = args.rotation
+
+    scenes = os.listdir(path)
+    # remove min max file from scenes folder
+    scenes = [s for s in scenes if min_max_filename not in s]
+
+    # getting image zone size and usefull information
+    zone_width, zone_height = image_zone_size
+    scene_width, scene_height = image_scene_size
+    nb_x_parts = math.floor(scene_width / zone_width)
+
+    output_dataset_filename_path = os.path.join(p_output, data_augmented_filename)
+
+    # go ahead each scenes
+    for folder_scene in scenes:
+
+        scene_path = os.path.join(path, folder_scene)
+
+        # build output scene path
+        output_scene_path = os.path.join(p_output, folder_scene)
+
+        if not os.path.exists(output_scene_path):
+            os.makedirs(output_scene_path)
+
+        # construct each zones folder name
+        zones_folder = []
+        zones_threshold = []
+
+        # get zones list info
+        for index in zones:
+            index_str = str(index)
+            if len(index_str) < 2:
+                index_str = "0" + index_str
+
+            current_zone = "zone"+index_str
+            zones_folder.append(current_zone)
+
+            zone_path = os.path.join(scene_path, current_zone)
+
+            with open(os.path.join(zone_path, seuil_expe_filename)) as f:
+                zones_threshold.append(int(f.readline()))
+
+        possible_x, possible_y = possible_point_zone
+
+        # get all images of folder
+        scene_images = sorted([os.path.join(scene_path, img) for img in os.listdir(scene_path) if cfg.scene_image_extension in img])
+        number_scene_image = len(scene_images)
+
+        for id_img, img_path in enumerate(scene_images):
+            
+            current_img = Image.open(img_path)
+            img = np.array(current_img)
+
+            for generation in range(p_number):
+                p_x, p_y = (random.randrange(possible_x), random.randrange(possible_y))
+
+                # extract random zone into scene image
+                extracted_img = img[p_y:(p_y + zone_height), p_x:(p_x + zone_width)]
+                
+                extracted_img.shape
+
+                pil_extracted_img = Image.fromarray(extracted_img)
+
+                # coordinate of specific zone, hence use threshold of zone
+                if p_x % zone_width == 0 and p_y % zone_height == 0:
+                    
+                    zone_index = math.floor(p_x / zone_width) + math.floor(p_y / zone_height) * nb_x_parts
+
+                    final_threshold = int(zones_threshold[zone_index])
+                else:
+                    # get zone identifiers of this new zones (from endpoints)
+                    p_top_left = (p_x, p_y)
+                    p_top_right = (p_x + zone_width, p_y)
+                    p_bottom_right = (p_x + zone_width, p_y + zone_height)
+                    p_bottom_left = (p_x, p_y + zone_height)
+
+                    points = [p_top_left, p_top_right, p_bottom_right, p_bottom_left]
+
+                    p_zones_indices = []
+                    # for each points get threshold information
+                    for p in points:
+                        x, y = p
+
+                        zone_index = math.floor(x / zone_width) + math.floor(y / zone_height) * nb_x_parts
+                        p_zones_indices.append(zone_index)
+
+                    p_thresholds = np.array(zones_threshold)[p_zones_indices]
+
+                    # get proportions of pixels of img into each zone
+                    overlaps = []
+
+                    p_x_max = p_x + zone_width
+                    p_y_max = p_y + zone_height
+
+                    for index, zone_index in enumerate(p_zones_indices):
+                        x_zone = (zone_index % nb_x_parts) * zone_width
+                        y_zone = (math.floor(zone_index / nb_x_parts)) * zone_height
+
+                        x_max_zone = x_zone + zone_width
+                        y_max_zone = y_zone + zone_height
+
+                        # computation of overlap
+                        # x_overlap = max(0, min(rect1.right, rect2.right) - max(rect1.left, rect2.left))
+                        # y_overlap = max(0, min(rect1.bottom, rect2.bottom) - max(rect1.top, rect2.top))
+                        x_overlap = max(0, min(x_max_zone, p_x_max) - max(x_zone, p_x))
+                        y_overlap = max(0, min(y_max_zone, p_y_max) - max(y_zone, p_y))
+
+                        overlapArea = x_overlap * y_overlap
+                        overlaps.append(overlapArea)
+
+                    overlapSum = sum(overlaps)
+                    proportions = [item / overlapSum for item in overlaps]
+                    
+                    final_threshold = 0
+
+                    for index, proportion in enumerate(proportions):
+                        final_threshold += proportion * p_thresholds[index]
+                    
+                    final_threshold = int(final_threshold)
+
+                # save image into new scene folder
+                current_image_postfix = dt.get_scene_image_postfix(img_path)
+
+                # prepare output img name
+                label_img = (int(current_image_postfix) < final_threshold)
+                extracted_image_name = dt.get_scene_image_prefix(img_path) + '_' + str(generation) + '_x' + str(p_x) + '_y' + str(p_y) + '_label' + str(int(label_img))
+
+                # if wished add of rotations images with same final threshold (increase data)
+                # write new line into global .csv ('threshold', 'filepath')
+                if p_rotation:
+
+                    # do rotations and save
+                    rotations = [0, 90, 180, 270]
+
+                    for rotation in rotations:
+
+                        rotated_img_name = extracted_image_name +  'rot' + str(rotation) + '_' + current_image_postfix + cfg.scene_image_extension
+                        rotated_img_path = os.path.join(output_scene_path, rotated_img_name)
+                        rotated_img = pil_extracted_img.rotate(rotation)
+                        rotated_img.save(os.path.join(rotated_img_path))
+
+                        csv_line = folder_scene + ';' + str(final_threshold) + ';' + str(int(current_image_postfix)) + ';' + str(int(label_img)) + ';' + rotated_img_path + '\n'
+
+                        with open(output_dataset_filename_path, 'a') as f:
+                            f.write(csv_line)
+
+                else:
+                    extracted_image_name += current_image_postfix + cfg.scene_image_extension
+                    extracted_image_path = os.path.join(output_scene_path, extracted_image_name)
+                    
+                    pil_extracted_img.save(extracted_image_path)
+
+                    csv_line = folder_scene + ';' + str(final_threshold) + ';' + str(int(current_image_postfix)) + ';' + str(int(label_img)) + ';' + extracted_image_path + '\n'
+
+                    with open(output_dataset_filename_path, 'a') as f:
+                        f.write(csv_line)
+
+                print(folder_scene + " - " + "{0:.2f}".format(((id_img * p_number + generation) + 1) / (p_number * number_scene_image) * 100.) + "%")
+                sys.stdout.write("\033[F")
+
+
+if __name__== "__main__":
+    main()