Spaces:
Runtime error
Runtime error
Upload 8 files
Browse files- .gitignore +2 -0
- app.py +367 -0
- corners-best.pt +3 -0
- example1.jpg +0 -0
- example2.jpg +0 -0
- keypoints-best.pt +3 -0
- requirements.txt +8 -0
.gitignore
ADDED
@@ -0,0 +1,2 @@
|
|
|
|
|
|
|
1 |
+
gradio_cached_examples/
|
2 |
+
temp*.jpg
|
app.py
ADDED
@@ -0,0 +1,367 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
#
|
2 |
+
import gradio as gr
|
3 |
+
|
4 |
+
from ultralytics import YOLO
|
5 |
+
|
6 |
+
from google.cloud import vision
|
7 |
+
|
8 |
+
client = vision.ImageAnnotatorClient()
|
9 |
+
|
10 |
+
import math
|
11 |
+
from scipy.spatial import KDTree
|
12 |
+
import io
|
13 |
+
from time import time
|
14 |
+
from PIL import Image, ImageDraw
|
15 |
+
import numpy as np
|
16 |
+
import cv2
|
17 |
+
|
18 |
+
from typing import Union
|
19 |
+
|
20 |
+
modelPh = r'corners-best.pt'
|
21 |
+
|
22 |
+
model1DIM = 640
|
23 |
+
keypointModel = r'keypoints-best.pt'
|
24 |
+
|
25 |
+
|
26 |
+
_examples = ["example1.jpg", "example2.jpg"]
|
27 |
+
|
28 |
+
|
29 |
+
def unwarp_image(warped_image, src_points, dst_points, output_width, output_height):
|
30 |
+
src_pts = np.array(src_points).astype(np.float64)
|
31 |
+
dst_pts = np.array(dst_points).astype(np.float64)
|
32 |
+
|
33 |
+
homography, mask = cv2.findHomography(src_pts, dst_pts)
|
34 |
+
unwarped_image = cv2.warpPerspective(
|
35 |
+
np.array(warped_image), homography, (output_width, output_height)
|
36 |
+
)
|
37 |
+
unwarped_image = Image.fromarray(unwarped_image)
|
38 |
+
|
39 |
+
return unwarped_image
|
40 |
+
|
41 |
+
|
42 |
+
model0 = None
|
43 |
+
|
44 |
+
def get_load_PhModel():
|
45 |
+
global model0
|
46 |
+
if model0 ==None:
|
47 |
+
tic = time()
|
48 |
+
model0 = YOLO(modelPh) # load a custom model
|
49 |
+
print(f"model0 load took: {time()-tic:.2g}")
|
50 |
+
return model0
|
51 |
+
|
52 |
+
|
53 |
+
def get_corners(results:list, img):
|
54 |
+
global model1DIM
|
55 |
+
# keypoints ie corners for homography
|
56 |
+
KP = "topLeft topRight bottomRight bottomLeft".split()
|
57 |
+
r = results[0]
|
58 |
+
kpco = r.keypoints.xy.cpu().squeeze()
|
59 |
+
assert len(kpco)>0, "not found"
|
60 |
+
keypoints = {k:v.numpy() for v,k in zip(kpco,KP)}
|
61 |
+
sz = model1DIM
|
62 |
+
dstCorners = np.array([(0,0),(sz,0),(sz,sz),(0,sz)])
|
63 |
+
planar = unwarp_image(img, np.array(list(keypoints.values())),dstCorners, sz,sz)
|
64 |
+
# planar.save("temp-ph.jpg")
|
65 |
+
return planar, keypoints
|
66 |
+
|
67 |
+
|
68 |
+
model = None
|
69 |
+
|
70 |
+
def get_load_KpModel():
|
71 |
+
global model
|
72 |
+
if model == None:
|
73 |
+
tic = time()
|
74 |
+
model = YOLO(keypointModel) # load a custom model
|
75 |
+
print(f"model load took: {time()-tic:.2g}")
|
76 |
+
return model
|
77 |
+
|
78 |
+
def preprocessImg(planar):
|
79 |
+
img = planar.convert('RGB').copy()
|
80 |
+
w,h = img.size
|
81 |
+
smalldl = abs(w-h)/h <0.05
|
82 |
+
_ = max(w,h)
|
83 |
+
DIM = w
|
84 |
+
if w!=h and smalldl:
|
85 |
+
img = img.resize((_,_))
|
86 |
+
elif w!=h:
|
87 |
+
img = img.resize((_,_))
|
88 |
+
|
89 |
+
return img
|
90 |
+
|
91 |
+
|
92 |
+
def get_keypoints(results:list):
|
93 |
+
assert len(results) ==1, "found multiple dials. expected only 1"
|
94 |
+
r = results[0]
|
95 |
+
# ordering
|
96 |
+
kp = "start_kp center end_kp tip".split()
|
97 |
+
kpco = r.keypoints.xy.cpu().squeeze()
|
98 |
+
keypoints = {k:v.numpy() for v,k in zip(kpco,kp)}
|
99 |
+
assert len(keypoints["center"])==2, "center keypoint not found"
|
100 |
+
assert len(keypoints["tip"])==2, "tip keypoint not found"
|
101 |
+
return keypoints
|
102 |
+
|
103 |
+
def cosangle(a,b, ignoreRot=False):
|
104 |
+
na = np.linalg.norm(a)
|
105 |
+
nb = np.linalg.norm(b)
|
106 |
+
angle2tip = np.rad2deg(np.arccos(np.dot(a, b)/(na*nb)))
|
107 |
+
angle2tip
|
108 |
+
rotdir = np.cross(a,b) < 0
|
109 |
+
if rotdir and not ignoreRot:
|
110 |
+
return 360-angle2tip
|
111 |
+
return angle2tip
|
112 |
+
|
113 |
+
|
114 |
+
def calculate_sweep_angles(keypoints:dict):
|
115 |
+
# get sweep angles start->tip
|
116 |
+
a = keypoints["start_kp"] - keypoints["center"]
|
117 |
+
b = keypoints["tip"] - keypoints["center"]
|
118 |
+
angle2tip = cosangle(a, b)
|
119 |
+
# get sweep angles start->end
|
120 |
+
b = keypoints["end_kp"] - keypoints["center"]
|
121 |
+
totalAngle = cosangle(a, b)
|
122 |
+
return angle2tip, totalAngle
|
123 |
+
|
124 |
+
|
125 |
+
def get_text_from_image(client, path_or_img)->Union[list[dict],Exception ]:
|
126 |
+
if type(path_or_img)==str:
|
127 |
+
with open(path_or_img, "rb") as image_file:
|
128 |
+
content = image_file.read()
|
129 |
+
else:
|
130 |
+
buf = io.BytesIO()
|
131 |
+
path_or_img.save(buf, format="JPEG")
|
132 |
+
content = buf.getvalue()
|
133 |
+
|
134 |
+
image = vision.Image(content=content)
|
135 |
+
|
136 |
+
response = client.text_detection(image=image)
|
137 |
+
if response.error.message:
|
138 |
+
raise Exception(
|
139 |
+
"{}\nFor more info on error messages, check: "
|
140 |
+
"https://cloud.google.com/apis/design/errors".format(response.error.message)
|
141 |
+
)
|
142 |
+
|
143 |
+
texts = response.text_annotations
|
144 |
+
contents = [ {"text": found.description, "boxCorners": [ (vert.x, vert.y) for vert in found.bounding_poly.vertices]} for found in texts]
|
145 |
+
return contents
|
146 |
+
|
147 |
+
def median_point_of_bounding_box(x1, y1, x2, y2, x3, y3, x4, y4):
|
148 |
+
x_coords = [x1, x2, x3, x4]
|
149 |
+
y_coords = [y1, y2, y3, y4]
|
150 |
+
x_median = sum(x_coords) / len(x_coords)
|
151 |
+
y_median = sum(y_coords) / len(y_coords)
|
152 |
+
return x_median, y_median
|
153 |
+
|
154 |
+
def to_numeric(text:str):
|
155 |
+
try:
|
156 |
+
return float(text)
|
157 |
+
except:
|
158 |
+
pass
|
159 |
+
return None
|
160 |
+
|
161 |
+
def result_as_validvalue(contents:list[dict])->tuple[list[dict], list[str]]:
|
162 |
+
# only valid values and sort min to max
|
163 |
+
valid = []
|
164 |
+
other = []
|
165 |
+
for f in contents:
|
166 |
+
t = f["text"]
|
167 |
+
value = to_numeric(t)
|
168 |
+
if "\n" in t:
|
169 |
+
continue
|
170 |
+
elif value == None and t!="":
|
171 |
+
other.append(t)
|
172 |
+
continue
|
173 |
+
b = f["boxCorners"]
|
174 |
+
m = median_point_of_bounding_box(*np.array(b).flatten())
|
175 |
+
valid.append({"text":f["text"], "value": value, "mid": m})
|
176 |
+
|
177 |
+
valid.sort(key=lambda e: e["value"])
|
178 |
+
return valid, other
|
179 |
+
|
180 |
+
|
181 |
+
distance = lambda a,b : np.sqrt(np.square(np.array(a)-np.array(b)).sum())
|
182 |
+
|
183 |
+
def determine_ocr_neighbors(center, valid:list[dict])->tuple[ list, float ]:
|
184 |
+
def cosangle(a,b):
|
185 |
+
na = np.linalg.norm(a)
|
186 |
+
nb = np.linalg.norm(b)
|
187 |
+
ang = np.rad2deg(np.arccos(np.dot(a, b)/(na*nb)))
|
188 |
+
rotdir = -1 if np.cross(a,b) < 0 else 1
|
189 |
+
return ang , rotdir
|
190 |
+
# compute angles between values
|
191 |
+
values = [valid[0]]
|
192 |
+
values[0]["dang"] = 0
|
193 |
+
rates = []
|
194 |
+
angS = 0
|
195 |
+
for v in valid[1:]:
|
196 |
+
u = v.copy()
|
197 |
+
u["dv"] = v["value"] - values[-1]["value"]
|
198 |
+
a = np.array(values[-1]["mid"]) - center
|
199 |
+
b = np.array(v["mid"]) - center
|
200 |
+
ang,_ = cosangle(a,b)
|
201 |
+
if _ <0:
|
202 |
+
Warning(f"skipping {u['value']} rot:{_}")
|
203 |
+
continue
|
204 |
+
angS += ang
|
205 |
+
u["dang"] = ang
|
206 |
+
u["dvda"] = u["dv"] / ang
|
207 |
+
rates.append(u["dvda"])
|
208 |
+
values.append(u)
|
209 |
+
|
210 |
+
rates = np.array(rates)
|
211 |
+
meanAng = angS/len(valid)
|
212 |
+
if len(rates)>=6:
|
213 |
+
ix = np.bitwise_and(rates> np.quantile(rates, 0.05) , rates<np.quantile(rates, 0.95))
|
214 |
+
if not np.all(~ix):
|
215 |
+
rates = rates[ix]
|
216 |
+
rate = rates.mean()
|
217 |
+
rate, meanAng
|
218 |
+
return values, rate
|
219 |
+
|
220 |
+
|
221 |
+
def vec_angle(v1, v2)->tuple[float, bool]:
|
222 |
+
vector1 = v1/np.linalg.norm(v1)
|
223 |
+
vector2 = v2/np.linalg.norm(v2)
|
224 |
+
angle_rad = np.arctan2(np.cross(vector1, vector2), np.dot(vector1, vector2))
|
225 |
+
return math.degrees(angle_rad)
|
226 |
+
|
227 |
+
|
228 |
+
def angles_from_tip(keypoints, values, nearestIx):
|
229 |
+
center = keypoints["center"]
|
230 |
+
tip = keypoints["tip"] - center
|
231 |
+
v = values[nearestIx[0]]
|
232 |
+
a = v["mid"] - center
|
233 |
+
ang = vec_angle(a,tip)
|
234 |
+
cumsum = 0
|
235 |
+
for i in range(nearestIx[0],-1,-1):
|
236 |
+
values[i]["before"] = abs(ang)+cumsum
|
237 |
+
cumsum += values[i]["dang"]
|
238 |
+
|
239 |
+
v = values[nearestIx[1]]
|
240 |
+
a = v["mid"] - center
|
241 |
+
ang = vec_angle(a,tip)
|
242 |
+
values[nearestIx[1]]["dang"] = 0
|
243 |
+
cumsum = 0
|
244 |
+
for i in range(nearestIx[1], len(values)):
|
245 |
+
cumsum -= values[i]["dang"]
|
246 |
+
values[i]["before"] = -abs(ang)+cumsum
|
247 |
+
|
248 |
+
return values
|
249 |
+
|
250 |
+
|
251 |
+
def get_needle_value(img, keypoints):
|
252 |
+
|
253 |
+
tic2 = time()
|
254 |
+
contents = get_text_from_image(client, img)
|
255 |
+
toc = time()
|
256 |
+
print(f"ocr took: {toc-tic2:.1g}")
|
257 |
+
|
258 |
+
assert len(contents)
|
259 |
+
valid,other = result_as_validvalue(contents)
|
260 |
+
assert len(valid)
|
261 |
+
center = np.array(keypoints["center"])
|
262 |
+
values, rate = determine_ocr_neighbors(center, valid)
|
263 |
+
assert len(values)>=2, "failed to find at least 2 OCR values"
|
264 |
+
|
265 |
+
# import pandas as pd
|
266 |
+
# print(pd.DataFrame.from_dict(values))
|
267 |
+
|
268 |
+
tree = KDTree([v["mid"] for v in values])
|
269 |
+
# find bounding ocr values of tip
|
270 |
+
dist, nearestIx = tree.query(keypoints["tip"],k=2)
|
271 |
+
nearestIx.sort()
|
272 |
+
dist, nearestIx
|
273 |
+
|
274 |
+
values = angles_from_tip(keypoints, values, nearestIx)
|
275 |
+
# compare against start and end
|
276 |
+
c = keypoints["center"]
|
277 |
+
tip = keypoints["tip"] - c
|
278 |
+
|
279 |
+
tipValues = []
|
280 |
+
for i in range(len(values)):
|
281 |
+
v = values[i]
|
282 |
+
a = v["mid"] - c
|
283 |
+
ang = vec_angle(a,tip)
|
284 |
+
before = v["before"]
|
285 |
+
startValue = v["value"]
|
286 |
+
angle2tip = ang
|
287 |
+
needleVal = -1
|
288 |
+
|
289 |
+
angle2tip = before
|
290 |
+
|
291 |
+
needleVal = angle2tip * rate + startValue # tip value from nearest Ix
|
292 |
+
tipValues.append(needleVal)
|
293 |
+
print(f"{i}, {ang:.2f}, {before:.2f}, @{needleVal:.2f}, {startValue}")
|
294 |
+
|
295 |
+
# print(f"total took: {toc-tic:.1g}")
|
296 |
+
tipValues = np.array(tipValues)
|
297 |
+
|
298 |
+
debug(img, contents, keypoints)
|
299 |
+
|
300 |
+
startValue= float(values[0]["value"])
|
301 |
+
tipvalue= round(tipValues[nearestIx].mean(),2)
|
302 |
+
endValue= float(values[-1]["value"])
|
303 |
+
|
304 |
+
return {"startValue": startValue, "tipvalue": tipvalue, "endValue": endValue, "unitPerDeg": float(rate), "otherText": list(set(other))}
|
305 |
+
|
306 |
+
|
307 |
+
# debug draw
|
308 |
+
def corners2bbox(C):
|
309 |
+
p = np.array(C)
|
310 |
+
s,e = p.min(axis=0).astype(int), p.max(axis=0).astype(int)
|
311 |
+
return s, e
|
312 |
+
|
313 |
+
def debug(img, contents, keypoints):
|
314 |
+
draw = ImageDraw.Draw(img)
|
315 |
+
|
316 |
+
for f in contents:
|
317 |
+
b = f["boxCorners"]
|
318 |
+
s,e = corners2bbox(b)
|
319 |
+
c = (255,0,0)
|
320 |
+
draw.rectangle((*s,*e), fill=None, outline=c, width=1)
|
321 |
+
m = median_point_of_bounding_box(*np.array(b).flatten())
|
322 |
+
draw.point(m, (255,0,255))
|
323 |
+
img
|
324 |
+
|
325 |
+
for v,c in zip(keypoints.values(), [(255,0,0), (0,255,0), (0,0,255),(255,0,255)]):
|
326 |
+
s = np.array(v)-1
|
327 |
+
e = np.array(v)+1
|
328 |
+
draw.rectangle((*s,*e), c)
|
329 |
+
img.save("temp-ocr.jpg")
|
330 |
+
print("saved debug img")
|
331 |
+
|
332 |
+
|
333 |
+
def predict(img, detect_gauge_first):
|
334 |
+
if detect_gauge_first:
|
335 |
+
model0 = get_load_PhModel()
|
336 |
+
results = model0.predict(img)
|
337 |
+
phimg,_ = get_corners(results, img)
|
338 |
+
else:
|
339 |
+
phimg = img.copy()
|
340 |
+
|
341 |
+
model = get_load_KpModel()
|
342 |
+
phimg = preprocessImg(phimg)
|
343 |
+
results = model.predict(phimg)
|
344 |
+
keypoints = get_keypoints(results)
|
345 |
+
|
346 |
+
angle2tip, totalAngle = calculate_sweep_angles(keypoints)
|
347 |
+
|
348 |
+
payload = get_needle_value(phimg, keypoints)
|
349 |
+
payload["angleToTip"] = round(angle2tip,2)
|
350 |
+
payload["totalAngle"] = round(totalAngle,2)
|
351 |
+
|
352 |
+
return payload
|
353 |
+
|
354 |
+
|
355 |
+
def test(img, detect_gauge_first):
|
356 |
+
return {"msg":str(img.size), "other": detect_gauge_first}
|
357 |
+
|
358 |
+
|
359 |
+
gr.Interface(fn=predict,
|
360 |
+
inputs=[
|
361 |
+
gr.Image(type="pil", sources=["upload","clipboard"],streaming=False, min_width=640),
|
362 |
+
gr.Checkbox(True, label="detect gauge first", info="if input image is zoomed in on only one gauge, uncheck box")
|
363 |
+
],
|
364 |
+
outputs="json",
|
365 |
+
examples=[_examples],
|
366 |
+
cache_examples=True)\
|
367 |
+
.launch()
|
corners-best.pt
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:a88502e86a40941aec69fe4d48e03c675a9381500fbf4c1ca8e3d1a89db089a9
|
3 |
+
size 37732202
|
example1.jpg
ADDED
example2.jpg
ADDED
keypoints-best.pt
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:d583485a30cd58e986231e7a02b84ce86e117d7eb48d4b5a901e4bada55319ac
|
3 |
+
size 6408962
|
requirements.txt
ADDED
@@ -0,0 +1,8 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
ultralytics==8.1.2
|
2 |
+
opencv-python==4.9.0.80
|
3 |
+
opencv-python-headless==4.8.0.76
|
4 |
+
numpy==1.24.1
|
5 |
+
scipy==1.11.2
|
6 |
+
gradio_client==0.8.0
|
7 |
+
google-cloud-vision==3.5.0
|
8 |
+
Pillow==9.3.0
|