MBA98 commited on
Commit
a18ccbe
·
1 Parent(s): 446a989

Create app.py

Browse files
Files changed (1) hide show
  1. app.py +126 -0
app.py ADDED
@@ -0,0 +1,126 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import gradio as gr
2
+ import cv2
3
+ import numpy as np
4
+ import tensorflow as tf
5
+
6
+ IMAGE_SIZE = 128
7
+ NUM_CLASSES = 3
8
+
9
+ def preprocess_image(image, target_size, add_clahe=True, clip_limit=4, tile_grid_size=(40,40), all_clahe=True):
10
+ """
11
+ Preprocess the images to remove black borders
12
+ and improve contrast using Y channel and CLAHE
13
+ """
14
+ try:
15
+ # Crop the image to remove black borders
16
+ cropped_rgb_image = remove_black_borders(image)
17
+
18
+ if add_clahe:
19
+ equalised_image = apply_clahe(
20
+ cropped_rgb_image, clip_limit, tile_grid_size, all_clahe)
21
+
22
+ # Resize the image to target size
23
+ resized_image = cv2.resize(equalised_image, (target_size, target_size))
24
+ else:
25
+ # Resize the image to target size
26
+ resized_image = cv2.resize(cropped_rgb_image, (target_size, target_size))
27
+
28
+ return resized_image
29
+
30
+ except Exception as e:
31
+ print("Error processing image: ")
32
+ print(e)
33
+
34
+
35
+ def remove_black_borders(image):
36
+ """
37
+ Crop the image to remove black borders
38
+ """
39
+ green_channel_image = image[:, :, 1]
40
+
41
+ # Find the contours in the green channel
42
+ contours, _ = cv2.findContours(
43
+ green_channel_image, cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)
44
+
45
+ # Find the largest contour
46
+ largest_contour = max(contours, key=cv2.contourArea)
47
+
48
+ # Get the bounding rectangle of the largest contour
49
+ x, y, w, h = cv2.boundingRect(largest_contour)
50
+
51
+ # Create a mask with the same size as the bounding rectangle
52
+ mask = np.zeros((h, w), np.uint8)
53
+
54
+ # Draw the largest contour on the mask
55
+ cv2.drawContours(mask, [largest_contour - [x, y]], 0, 255, -1)
56
+
57
+ # Convert the mask to a 3 channel image
58
+ mask_3_channel = cv2.merge([mask, mask, mask])
59
+
60
+ # Crop the image using the mask
61
+ cropped_image = cv2.bitwise_and(image[y:y+h, x:x+w], mask_3_channel)
62
+
63
+ return cropped_image
64
+
65
+
66
+ def apply_clahe(image, clip_limit=4, tile_grid_size=(40,40), all_clahe=False):
67
+ """
68
+ Preprocess the image
69
+ """
70
+ try:
71
+ # Extract the y channel from the cropped rgb image
72
+ yuv_image = cv2.cvtColor(image, cv2.COLOR_RGB2YUV)
73
+ cropped_y_channel_image = yuv_image[:, :, 0]
74
+
75
+ # Apply CLAHE to the y channel of the cropped image
76
+ clahe = cv2.createCLAHE(clipLimit=clip_limit, tileGridSize=tile_grid_size)
77
+ enhanced_y_channel = clahe.apply(cropped_y_channel_image)
78
+
79
+ if all_clahe:
80
+ cropped_u_channel = yuv_image[:, :, 1]
81
+ enhanced_u_channel = clahe.apply(cropped_u_channel)
82
+ cropped_v_channel = yuv_image[:, :, 2]
83
+ enhanced_v_channel = clahe.apply(cropped_v_channel)
84
+
85
+ enhanced_yuv_image = np.stack([enhanced_y_channel, enhanced_u_channel, enhanced_v_channel], axis=-1)
86
+
87
+ # Convert YUV to RGB
88
+ enhanced_rgb_image = cv2.cvtColor(enhanced_yuv_image, cv2.COLOR_YUV2RGB)
89
+
90
+ return enhanced_rgb_image
91
+ else:
92
+ # Convert the y channel image to grayscale
93
+ enhanced_grayscale_image = cv2.convertScaleAbs(enhanced_y_channel, alpha=(255/219))
94
+
95
+ # Repeat the equalised grayscale for all 3-channels
96
+ enhanced_grayscale_3_channels = np.stack((enhanced_grayscale_image,) * 3, axis=-1)
97
+
98
+ return enhanced_grayscale_3_channels
99
+
100
+ except Exception as e:
101
+ print("Error processing image: ", e)
102
+
103
+
104
+ def load_vgg16_model():
105
+ vgg16_bs32_img128_CLAHE4 = tf.keras.models.load_model('./VGG16_32_128_CLAHE_4_40_trainableFalse.h5')
106
+ return vgg16_bs32_img128_CLAHE4
107
+
108
+
109
+ def predict_input_image(img):
110
+ # img = preprocess_image(img, IMAGE_SIZE, add_clahe=True, clip_limit=4, tile_grid_size=(40,40), all_clahe=True)
111
+ # img_4d=img.reshape(-1, IMAGE_SIZE, IMAGE_SIZE, 3)
112
+ # img_4d_ensemble = [img_4d for _ in range(len(ensemble_model.input))]
113
+ # prediction=ensemble_model.predict([img_4d_ensemble])[0]
114
+
115
+ img = preprocess_image(img, IMAGE_SIZE, add_clahe=True, clip_limit=4, tile_grid_size=(40,40), all_clahe=False)
116
+ img_4d=img.reshape(-1, IMAGE_SIZE, IMAGE_SIZE, 3)
117
+ model = load_vgg16_model()
118
+ prediction=model.predict([img_4d])[0]
119
+
120
+ dr_classes = ['No DR', 'Mild DR', 'Referable DR']
121
+ return {dr_classes[i]: float(prediction[i]) for i in range(3)}
122
+
123
+ image = gr.inputs.Image(shape=(IMAGE_SIZE, IMAGE_SIZE))
124
+ label = gr.outputs.Label(num_top_classes=NUM_CLASSES)
125
+ iface = gr.Interface(fn=predict_input_image, inputs=image, outputs=label).launch(share=True)
126
+ iface.launch()