Create app.py
Browse files
app.py
ADDED
@@ -0,0 +1,190 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import gradio as gr
|
2 |
+
import onnxruntime as ort
|
3 |
+
import numpy as np
|
4 |
+
from PIL import Image, ImageDraw
|
5 |
+
import cv2
|
6 |
+
|
7 |
+
image_size = 224
|
8 |
+
|
9 |
+
def normalize_image(image, mean=(0.485, 0.456, 0.406), std=(0.229, 0.224, 0.225)):
|
10 |
+
image = (image/255.0).astype("float32")
|
11 |
+
|
12 |
+
image[:, :, 0] = (image[:, :, 0] - mean[0]) / std[0]
|
13 |
+
image[:, :, 1] = (image[:, :, 1] - mean[1]) / std[1]
|
14 |
+
image[:, :, 2] = (image[:, :, 2] - mean[2]) / std[2]
|
15 |
+
|
16 |
+
return image
|
17 |
+
|
18 |
+
def resize_longest_max_size(image, max_size=224):
|
19 |
+
|
20 |
+
height, width = image.shape[:2]
|
21 |
+
|
22 |
+
if width > height:
|
23 |
+
ratio = max_size / width
|
24 |
+
else:
|
25 |
+
ratio = max_size / height
|
26 |
+
|
27 |
+
new_width = int(width * ratio)
|
28 |
+
new_height = int(height * ratio)
|
29 |
+
|
30 |
+
resized_image = cv2.resize(image, (new_width, new_height), cv2.INTER_LINEAR)
|
31 |
+
|
32 |
+
return resized_image
|
33 |
+
|
34 |
+
def pad_if_needed(image, target_size):
|
35 |
+
height, width, _ = image.shape
|
36 |
+
|
37 |
+
y0 = abs((height-target_size)//2)
|
38 |
+
x0 = abs((width-target_size)//2)
|
39 |
+
|
40 |
+
background = np.zeros((target_size, target_size, 3), dtype="uint8")
|
41 |
+
background[y0:(y0+height), x0:(x0+width), :] = image
|
42 |
+
|
43 |
+
return(background)
|
44 |
+
|
45 |
+
def heatmap2keypoints(heatmap: np.ndarray, image_size: int = 224) -> list:
|
46 |
+
"Function to convert heatmap to keypoint x, y tensor"
|
47 |
+
|
48 |
+
indx = heatmap.reshape(-1, image_size*image_size).argmax(axis=1)
|
49 |
+
row = indx // image_size
|
50 |
+
col = indx % image_size
|
51 |
+
|
52 |
+
keypoints_array = np.stack((col, row), axis=1)
|
53 |
+
keypoints_list = keypoints_array.tolist()
|
54 |
+
|
55 |
+
return keypoints_list
|
56 |
+
|
57 |
+
def centercrop_keypoints(keypoints, crop_height, crop_width, image_height, image_width):
|
58 |
+
y_diff = (image_height-crop_height)//2
|
59 |
+
x_diff = (image_width-crop_width)//2
|
60 |
+
|
61 |
+
keypoints_crop = [[x-x_diff, y-y_diff] for x, y in keypoints]
|
62 |
+
return(keypoints_crop)
|
63 |
+
|
64 |
+
def resize_keypoints(keypoints, current_height, current_width, target_height, target_width):
|
65 |
+
keypoints_resize = []
|
66 |
+
for x, y in keypoints:
|
67 |
+
x_resize = (x/current_width)*target_width
|
68 |
+
y_resize = (y/current_height)*target_height
|
69 |
+
keypoints_resize.append([int(x_resize), int(y_resize)])
|
70 |
+
return(keypoints_resize)
|
71 |
+
|
72 |
+
def draw_keypoints(image, keypoints):
|
73 |
+
draw = ImageDraw.Draw(image)
|
74 |
+
w, h = image.size
|
75 |
+
for keypoint in keypoints:
|
76 |
+
x, y = keypoint
|
77 |
+
# Draw a small circle at each keypoint
|
78 |
+
radius = int(min(w, h) * 0.01)
|
79 |
+
draw.ellipse((x - radius, y - radius, x + radius, y + radius), fill='red', outline='red')
|
80 |
+
return image
|
81 |
+
|
82 |
+
def point_dist(p0, p1):
|
83 |
+
x0, y0 = p0
|
84 |
+
x1, y1 = p1
|
85 |
+
|
86 |
+
dist = ((x0-x1)**2 + (y0-y1)**2)**0.5
|
87 |
+
|
88 |
+
return dist
|
89 |
+
|
90 |
+
def receipt_asp_ratio(keypoints, mode = "mean"):
|
91 |
+
|
92 |
+
h0 = point_dist(keypoints[0], keypoints[3])
|
93 |
+
h1 = point_dist(keypoints[1], keypoints[2])
|
94 |
+
|
95 |
+
w0 = point_dist(keypoints[0], keypoints[1])
|
96 |
+
w1 = point_dist(keypoints[2], keypoints[3])
|
97 |
+
|
98 |
+
if mode == "max":
|
99 |
+
h = max(h0, h1)
|
100 |
+
w = max(w0, w1)
|
101 |
+
elif mode == "mean":
|
102 |
+
h = (h0+h1)/2
|
103 |
+
w = (w0+w1)/2
|
104 |
+
else:
|
105 |
+
return("UNKNOWN MODE")
|
106 |
+
|
107 |
+
return w/h
|
108 |
+
|
109 |
+
# Load the ONNX model
|
110 |
+
session = ort.InferenceSession("models/timm-mobilenetv3_small_100.onnx")
|
111 |
+
|
112 |
+
input_name = session.get_inputs()[0].name
|
113 |
+
output_name = session.get_outputs()[0].name
|
114 |
+
|
115 |
+
# Main function to handle the image input, apply preprocessing, run the model, and apply postprocessing
|
116 |
+
def process_image(input_image):
|
117 |
+
|
118 |
+
# Convert PIL image to OpenCV image
|
119 |
+
image = np.array(input_image.convert("RGB"))
|
120 |
+
|
121 |
+
h, w, _ = image.shape
|
122 |
+
|
123 |
+
# Preprocess the image
|
124 |
+
image_resize = resize_longest_max_size(image)
|
125 |
+
|
126 |
+
h_small, w_small, _ = image_resize.shape
|
127 |
+
|
128 |
+
image_pad = pad_if_needed(image_resize, target_size=image_size)
|
129 |
+
|
130 |
+
image_norm = normalize_image(image_pad)
|
131 |
+
|
132 |
+
image_array = np.transpose(image_norm, (2, 0, 1))
|
133 |
+
|
134 |
+
image_array = np.expand_dims(image_array, axis=0)
|
135 |
+
|
136 |
+
# Run model inference
|
137 |
+
output = session.run([output_name], {input_name: image_array})
|
138 |
+
|
139 |
+
output_keypoints = heatmap2keypoints(output[0].squeeze())
|
140 |
+
crop_keypoints = centercrop_keypoints(output_keypoints, h_small, w_small, image_size, image_size)
|
141 |
+
large_keypoints = resize_keypoints(crop_keypoints, h_small, w_small, h, w)
|
142 |
+
|
143 |
+
# Draw keypoints on the image
|
144 |
+
image_with_keypoints = draw_keypoints(input_image, large_keypoints)
|
145 |
+
|
146 |
+
persp_h = 1024
|
147 |
+
persp_asp = receipt_asp_ratio(large_keypoints, mode="max")
|
148 |
+
persp_w = int(persp_asp*persp_h)
|
149 |
+
|
150 |
+
origin_points = np.float32([[x, y] for x, y in large_keypoints])
|
151 |
+
target_points = np.float32([[0, 0], [persp_w-1, 0], [persp_w-1, persp_h-1], [0, persp_h-1]])
|
152 |
+
|
153 |
+
persp_matrix = cv2.getPerspectiveTransform(origin_points, target_points)
|
154 |
+
persp_image = cv2.warpPerspective(image, persp_matrix, (persp_w, persp_h), cv2.INTER_LINEAR)
|
155 |
+
|
156 |
+
output_image = Image.fromarray(persp_image)
|
157 |
+
|
158 |
+
return image_with_keypoints, output_image
|
159 |
+
|
160 |
+
demo_images = [
|
161 |
+
"demo_images/image_1.jpg",
|
162 |
+
"demo_images/image_2.jpg",
|
163 |
+
"demo_images/image_3.jpg",
|
164 |
+
"demo_images/image_flux_1.png",
|
165 |
+
"demo_images/image_flux_2.png",
|
166 |
+
]
|
167 |
+
|
168 |
+
# Create Gradio interface
|
169 |
+
with gr.Blocks() as iface:
|
170 |
+
gr.Markdown("# Document corner detection and perspective correction")
|
171 |
+
gr.Markdown("Upload an image to detect the corners of a document and correct the perspective.\n\nUses a UNet model to detect corners and OpenCV to correct the perspective.")
|
172 |
+
|
173 |
+
with gr.Row():
|
174 |
+
with gr.Column():
|
175 |
+
input_image = gr.Image(type="pil", label="Image", show_label=True, scale=1)
|
176 |
+
|
177 |
+
with gr.Column():
|
178 |
+
output_image1 = gr.Image(type="pil", label="Image with predicted corners", show_label=True, scale=1)
|
179 |
+
|
180 |
+
with gr.Column():
|
181 |
+
output_image2 = gr.Image(type="pil", label="Image with perspective correction", show_label=True, scale=1)
|
182 |
+
|
183 |
+
with gr.Row():
|
184 |
+
examples = gr.Examples(demo_images, input_image, cache_examples=False, label="Exampled documents (CORD dataset and FLUX.1-schnell generated)")
|
185 |
+
|
186 |
+
input_image.change(fn=process_image, inputs=input_image, outputs=[output_image1, output_image2])
|
187 |
+
|
188 |
+
gr.Markdown("Created by Kenneth Thorø Martinsen (kenneth2810@gmail.com)")
|
189 |
+
|
190 |
+
iface.launch()
|