def main():
max_width()
config, regions, mmc, sdk_url = sidebar()
my_file = st.file_uploader('Pick an image',
type=['jpg', 'png', 'jpeg', 'gif', 'bmp'])
if not my_file:
return
image = Image.open(my_file).convert('RGB')
res = recognition(my_file, regions, sdk_url, mmc, config)
res = copy.deepcopy(res)
vehicles = []
for result in res['results']:
if 'vehicle' in result:
details = []
model_make = result.get('model_make')
if model_make:
details.append(
'{make} {model} {score:.2f}'.format(**model_make[0]))
color = result.get('color')
if color:
details.append('{color} {score:.2f}'.format(**color[0]))
result['vehicle']['details'] = ', '.join(details)
vehicles.append(result['vehicle'])
overlay = st.checkbox('Show Overlay', value=True)
if overlay:
draw_bb(
image, vehicles, None, lambda vehicle: '%s (%s) [%s]' %
(vehicle['type'], vehicle['score'], vehicle['details']))
draw_bb(
image, res['results'], None, lambda result: '%s [%s] (%s, %s)' %
(result['plate'].upper(), result['region']['code'], result[
'score'], result['dscore']))
st.image(image, width=image.width)
res
def process_image(path, args, i):
config = dict(threshold_d=args.detection_threshold,
threshold_o=args.ocr_threshold,
mode='redaction')
# Predictions
source_im = Image.open(path)
if source_im.mode != 'RGB':
source_im = source_im.convert('RGB')
images = [((0, 0), source_im)] # Entire image
# Top left and top right crops
if args.split_image:
y = 0
win_size = .55
width, height = source_im.width * win_size, source_im.height * win_size
for x in [0, int((1 - win_size) * source_im.width)]:
images.append(((x, y), source_im.crop(
(x, y, x + width, y + height))))
# Inference
results = []
for (x, y), im in images:
im_bytes = io.BytesIO()
im.save(im_bytes, 'JPEG', quality=95)
im_bytes.seek(0)
im_results = recognition_api(im_bytes,
args.regions,
args.api_key,
args.sdk_url,
config=config)
results.append(dict(prediction=im_results, x=x, y=y))
results = post_processing(merge_results(results))
results['filename'] = Path(path).name
# Set bounding box padding
for item in results['results']:
# Decrease padding size for large bounding boxes
b = item['box']
width, height = b['xmax'] - b['xmin'], b['ymax'] - b['ymin']
padding_x = int(
max(0, width * (.3 * math.exp(-10 * width / source_im.width))))
padding_y = int(
max(0, height * (.3 * math.exp(-10 * height / source_im.height))))
b['xmin'] = b['xmin'] - padding_x
b['ymin'] = b['ymin'] - padding_y
b['xmax'] = b['xmax'] + padding_x
b['ymax'] = b['ymax'] + padding_y
if args.show_boxes or args.save_blurred:
im = blur(source_im, 5, results)
if args.show_boxes:
im.show()
if args.save_blurred:
filename = Path(path)
im.save(filename.parent / ('%s_blurred%s' %
(filename.stem, filename.suffix)))
if 0:
draw_bb(source_im, results['results']).show()
return results
def process_image(path, args):
options = CAMERA_OPTIONS[args.camera]
config = dict(threshold_d=.2, threshold_o=.3, mode='redaction')
# Pre process image
source_im = Image.open(path)
rotated = source_im.rotate(options['rotation'], expand=True)
top = rotated.height * options['crop_top']
bottom = rotated.height * options['crop_bottom']
api_im = rotated.crop((0, top, rotated.width, bottom))
# Predictions
images = [((0, 0), api_im)] # Entire image
# Top left and top right crops
y = 0
win_size = .55
width, height = api_im.width * win_size, api_im.height * win_size
for x in [0, (1 - win_size) * api_im.width]:
images.append(((x, y), api_im.crop((x, y, x + width, y + height))))
# Inference
results = []
for (x, y), im in images:
im_bytes = io.BytesIO()
im.save(im_bytes, 'JPEG', quality=95)
im_bytes.seek(0)
im_results = recognition_api(im_bytes,
args.regions,
args.api_key,
args.sdk_url,
config=config)
results.append(dict(prediction=im_results, x=x, y=y))
results = merge_results(results)
# Update bounding boxes
for lp in results['results']:
box = lp['box']
box['ymin'] += top
box['ymax'] += top
# Rotate
lp['box'] = rotate_bb(box, options['rotation'], rotated.size)
if 0:
draw_bb(source_im, results['results']).show()
return results
def process_image(path, args, i):
config = dict(threshold_d=args.detection_threshold,
threshold_o=args.ocr_threshold,
mode='redaction')
# Predictions
source_im = Image.open(path)
images = [((0, 0), source_im)] # Entire image
# Top left and top right crops
if args.split_image:
y = 0
win_size = .55
width, height = source_im.width * win_size, source_im.height * win_size
for x in [0, int((1 - win_size) * source_im.width)]:
images.append(((x, y), source_im.crop(
(x, y, x + width, y + height))))
# Inference
results = []
for (x, y), im in images:
im_bytes = io.BytesIO()
if im.mode == 'RGBA':
im = im.convert('RGB')
im.save(im_bytes, 'JPEG', quality=95)
im_bytes.seek(0)
im_results = recognition_api(im_bytes,
args.regions,
args.api_key,
args.sdk_url,
config=config)
results.append(dict(prediction=im_results, x=x, y=y))
results = merge_results(results)
if args.show_boxes:
blur(source_im, 5, results).show()
if 0:
draw_bb(source_im, results['results']).show()
return results