I am running a weed detection pipeline which involves creating patches and running ML algorithm to detect region of weeds. When I try to now reconstruct the patches to create the complete weed map, the patches seem to be in disorder when recombined, see attached image .
here is the code that create patches:
class Patches:
def __init__(self, im_list, msk_list, PATCH_SIZE, threshold=0.03):
self.im_list = im_list
self.msk_list = msk_list
self.PATCH_SIZE = PATCH_SIZE
self.threshold = threshold
def image_to_patches(self, image, b_msk=False):
slc_size = self.PATCH_SIZE
x = int(np.ceil(image.shape[0] / slc_size))
y = int(np.ceil(image.shape[1] / slc_size))
padded_shape = (x * slc_size, y * slc_size)
if not b_msk:
padded_rgb_image = np.zeros((padded_shape[0], padded_shape[1], 3), dtype=np.uint8)
padded_rgb_image[:image.shape[0], :image.shape[1]] = image
patches = patchify(padded_rgb_image, (slc_size, slc_size, 3), step=slc_size)
else:
padded_rgb_image = np.zeros((padded_shape[0], padded_shape[1]), dtype=np.uint8)
padded_rgb_image[:image.shape[0], :image.shape[1]] = image
patches = patchify(padded_rgb_image, (slc_size, slc_size), step=slc_size)
return patches, slc_size
def load_image(self, path):
"""
loads an image based on the path
"""
rgb_image = skio.imread(path)
return rgb_image
def patchify_image_mask(self):
imgs = []
anns = []
AREA = self.PATCH_SIZE * self.PATCH_SIZE
f_AREA = int(self.threshold * AREA)
print(f'Threshold: {self.threshold} * {AREA} = {f_AREA}')
print(f"Patchyfying images and mask...")
for im_path, msk_path in zip(self.im_list, self.msk_list):
patches, _ = self.image_to_patches(self.load_image(im_path))
masks, _ = self.image_to_patches(self.load_image(msk_path), b_msk=True)
for i in range(patches.shape[0]):
for j in range(patches.shape[1]):
patch = patches[i, j, :, :, :]
mask = masks[i, j, ::]
if mask.reshape(-1).sum() > f_AREA:
imgs.append(patch)
anns.append(mask)
return np.array(imgs), np.array(anns)
and here is the test.py logic:
import sys
sys.path.append("/Users/i/Downloads/Deep-Weed-Segmentation-main/")
import numpy as np
import random
import matplotlib.pyplot as plt
from scripts.prepare_dataset import Prepare_Dataset
print("hallala")
import argparse
from scripts.model import Models
from tensorflow import keras
import segmentation_models as sm
import tensorflow_advanced_segmentation_models as tasm
import os
import cv2
from patchify import patchify, unpatchify
# /Users/icom/Downloads/CoFly-WeedDB
def main():
parser = argparse.ArgumentParser(description='Testing Model')
parser.add_argument('--network', help='Network Model type', default='custom')
parser.add_argument('--backbone', help='Backbone', default='None')
parser.add_argument('--weight_path', help='path to saved weights', default='./models')
parser.add_argument('--image_path', help='path to image', default='./images')
parser.add_argument('--patch_size', type=int, required=True, help='Size of the patches')
parser.add_argument('--binary', help='Enable Binary Segmentation', default=False, action='store_true')
parser.add_argument('--data_path', help='root path to dataset',
default='/Users/i/Downloads/CoFly-WeedDB')
args = parser.parse_args()
args = vars(args)
args = {k: v for k, v in args.items() if v is not None}
patch_size = int(args['patch_size'])
# Get the dataset and original dimensions
Y_train_cat, Y_test_cat, X_train, Y_test, X_test, p_weights, n_classes, original_height, original_width = Prepare_Dataset(
patch_size, binary=args['binary'], backbone=args['backbone'], data_path=args['data_path']
).prepare_all()
Test(args['network'], args['backbone'], args['weight_path'], args['image_path'], patch_size,
args['data_path'], args['binary']).test()
class Test:
def __init__(self, network, backbone, weight_path, input_image, PATCH_SIZE, data_path, binary):
selfwork = network
self.backbone = backbone
self.PATCH_SIZE = self.size_(PATCH_SIZE)
self.weights_path = weight_path
self.input_image = input_image
(self.Y_train_cat, self.Y_test_cat, self.X_train, self.Y_test, self.X_test, self.p_weights,
self.n_classes, self.original_height, self.original_width) = Prepare_Dataset(self.PATCH_SIZE, binary=binary, backbone=backbone, data_path=data_path).prepare_all()
self.total_loss = Prepare_Dataset(self.PATCH_SIZE).get_loss(p_weights=self.p_weights)
self.binary = binary
def size_(self, PATCH_SIZE):
if selfwork == 'pspnet':
if PATCH_SIZE % 48 != 0:
print('Image size must be divisible by 48')
PATCH_SIZE = int(PATCH_SIZE / 48) * 48 + 48
print(f'New image size: {PATCH_SIZE}x{PATCH_SIZE}x3')
return PATCH_SIZE
def test(self):
print('Testing...')
metrics = [sm.metrics.IOUScore(threshold=0.5), sm.metrics.FScore(threshold=0.5)]
LR = 0.0001
optim = keras.optimizers.Adam(LR)
print('Building Model...')
# Initialize model
if selfwork == 'custom':
model = Models(self.n_classes, self.PATCH_SIZE, IMG_CHANNELS=3, model_name=selfwork,
backbone=self.backbone).simple_unet_model()
elif selfwork == 'segnet':
model, self.backbone = Models(self.n_classes, self.PATCH_SIZE, IMG_CHANNELS=3, model_name=selfwork,
backbone=self.backbone).segnet_architecture()
elif selfwork == 'unet' or selfwork == 'linknet' or selfwork == 'pspnet':
model = Models(self.n_classes, self.PATCH_SIZE, IMG_CHANNELS=3, model_name=selfwork,
backbone=self.backbone).segmented_models()
elif selfwork == 'deeplabv3':
base_model, layers, layer_names = Models(self.n_classes, self.PATCH_SIZE, IMG_CHANNELS=3,
model_name=selfwork,
backbone=self.backbone).deeplabv3(name=self.backbone,
weights='imagenet',
height=self.PATCH_SIZE,
width=self.PATCH_SIZE)
model = tasm.DeepLabV3plus(n_classes=self.n_classes, base_model=base_model, output_layers=layers,
backbone_trainable=False)
model.build((None, self.PATCH_SIZE, self.PATCH_SIZE, 3))
else:
print(f'{selfwork} network not available.')
quit()
# Compilation
modelpile(optimizer=optim,
loss=self.total_loss,
metrics=[metrics])
print(model.summary())
path_to_model = os.path.join(self.weights_path,
selfwork + '_' + self.backbone + '_' + str(
self.PATCH_SIZE) + '_binary_' + str(self.binary) +
'.weights.h5')
try:
model.load_weights(path_to_model)
except Exception as err:
print(err)
quit()
# Initialize an empty array for the reconstructed original image with overlays
reconstructed_image_with_overlays = np.zeros((self.original_height, self.original_width, 3), dtype=np.uint8)
counter = 0
for img in range(len(self.X_test)):
counter += 1
test_img = self.X_test[img]
ground_truth = self.Y_test[img]
test_img_input = np.expand_dims(test_img, 0)
prediction = model.predict(test_img_input)
predicted_img = np.argmax(prediction, axis=3)[0, :, :]
# Create an overlay for the predicted areas
overlay = test_img.copy()
for class_id in range(1, self.n_classes): # Assuming class 0 is background
mask = predicted_img == class_id
overlay[mask] = [255, 0, 0] # Example: Red overlay for predicted areas
# Blend the overlay with the original test image
blended_overlay = cv2.addWeighted(test_img, 0.7, overlay, 0.3, 0)
# Save individual patch prediction
plt.figure(figsize=(16, 8))
plt.subplot(241)
plt.title('Testing Image')
plt.imshow(test_img)
plt.subplot(242)
plt.title('Testing Label')
plt.imshow(ground_truth, cmap='jet')
plt.subplot(243)
plt.title('Prediction on test image')
plt.imshow(predicted_img, cmap='jet')
plt.subplot(244)
plt.title('Overlay of Test Image and Prediction')
plt.imshow(blended_overlay)
plt.savefig(f'./plots/figure_{counter}_.png')
plt.close() # Close the figure to avoid memory issues
print(f'{(counter/len(self.X_test))*100}% done.')
# Calculate the starting row and column based on the patch index
num_patches_per_row = self.original_width // self.PATCH_SIZE
start_row = (img // num_patches_per_row) * self.PATCH_SIZE
start_col = (img % num_patches_per_row) * self.PATCH_SIZE
# Place the blended overlay patch in the reconstructed image
if start_row + self.PATCH_SIZE <= self.original_height and start_col + self.PATCH_SIZE <= self.original_width:
reconstructed_image_with_overlays[start_row:start_row + self.PATCH_SIZE, start_col:start_col + self.PATCH_SIZE] = blended_overlay
# Save the reconstructed original image with overlays
plt.figure(figsize=(12, 8))
plt.title('Reconstructed Original Image with Overlays')
plt.imshow(reconstructed_image_with_overlays)
plt.axis('off')
plt.savefig('./plots/reconstructed_image_with_overlays.png')
I am wondering why the patches are being reconstructed without order here, and what could be a suggestion?
Update: here is the input image that I was processing.