def test_resize_crop(self):
vertical_cropped = images.resize_crop(vertical, height, width)
self.assertEqual(vertical_cropped.shape, (height, height, 3))
horizontal_cropped = images.resize_crop(horizontal, height, width)
self.assertEqual(horizontal_cropped.shape, (height, height, 3))
square_cropped = images.resize_crop(square, height, width)
self.assertEqual(square_cropped.shape, (height, height, 3))
def results():
'''
Add image to database, calculates features and finds nearest neighbors.
:return: rendered template from html file
'''
# get data from form
search_file = request.files['search_file']
search_url = request.form['search_url']
k = int(request.form['count'])
# path to temporary image file
filename = 'img'
path = os.path.join(app.config['UPLOAD_FOLDER'], filename)
# list of images to be passed to render_template
filenames = []
with lock:
if not search_url and not search_file:
msg = 'Please provide URL or file.'
return render_template('results.html', msg=msg, data=[])
if search_url:
msg = 'Searching using URL (filename: %s).' % search_url.split('/')[-1]
# temporarily save image in full resolution
urllib.urlretrieve(search_url, path)
else:
msg = 'Searching using file (filename: %s).' % search_file.filename
# temporarily save image in full resolution
search_file.save(path)
# open image using opencv, resize it and crop it (BGR, float32)
img = cv2.imread(path)
if img is None:
msg = 'We apologize for the inconvenience but the image could not be saved.'
return render_template('results.html', msg=msg, filenames=filenames)
img_cropped = images.resize_crop(img, height, width)
img_cropped = img_cropped.astype(np.float32)
# delete temporary image
os.remove(path)
# calculate features (or choose random for local testing)
if caffe_toggle:
print 'Calculating features...'
score = net.predict([img_cropped]).flatten()
print 'Calculated.'
else:
with h5py.File(h5_fts_fn, 'r') as fr_features:
score = fr_features['score'][randint(0, 9999)]
# convert image to RGB uint8
img_rgb = cv2.cvtColor(img_cropped, cv2.COLOR_BGR2RGB)
img_rgb = img_rgb.astype(np.uint8)
# add image and save thumbnail
kdt.add_images([img_rgb], [score])
# find k nearest neighbors
last = len(kdt.features) - 1
indexes, distances = kdt.find_k_nearest_by_index(last, k+1)
# check for duplicate image
if (distances[1] - distances[0]) < 1e-10:
# remove duplicate image
kdt.remove_last_image()
# do not print duplicate image (recently added)
if indexes[0] > indexes[1]:
indexes = indexes[1:]
distances = distances[1:]
else:
del indexes[1]
del distances[1]
else:
indexes = indexes[:k]
distances = distances[:k]
kdt.save()
# prepare list of files which will be printed
for index in indexes:
filenames.append(str(index) + '.jpg')
zipped = zip(filenames, distances)
return render_template('results.html', msg=msg, data=zipped)
kdt = kdtree_flann.ImageSearchKDTreeFlann(storage_dir, 1000000000, (150, 150, 3))
i = 0
for path in imgs_paths:
if i % 50 == 0:
print '\r', i, path,
try:
# open image using opencv
img = cv2.imread(path)
# color channels
if len(img.shape) == 3:
# resize image and crop it (BGR, float32)
img_cropped = images.resize_crop(img, height, width)
img_cropped = img_cropped.astype(np.float32)
# calculate features
score = net.predict([img_cropped]).flatten()
# convert image to RGB uint8
img_rgb = cv2.cvtColor(img_cropped, cv2.COLOR_BGR2RGB)
img_rgb = img_rgb.astype(np.uint8)
# add image and save thumbnail
kdt.add_images([img_rgb], [score], build_tree=False)
i += 1
except:
print '\nNot an image'
