def evaluate():
img = Image.open(FLAGS.photo_file)
gray_img = img.convert('L')
img = [int(e) / 255.0 for e in resize(gray_img).tobytes()]
x = tf.placeholder(tf.float32, [None, 784])
kb = tf.placeholder(tf.float32, [])
convolution_op = convolution(x, kb)
with tf.Session() as sess:
saver = tf.train.Saver()
load_checkpoint(sess, saver)
logits = sess.run([convolution_op], feed_dict={x: [img], kb: 1.0})
print(LABEL2NAME[np.argmax(logits[0])])
def __init__(self, file_index, path, img_stats, height, width, multi,
whitelist_path):
NamedDB.__init__(self, whitelist_path, "FlipDB")
DB.__init__(self, "unaltered", img_stats, multi)
self.image_path = path
with open(file_index, "r") as f:
self.index = np.load(f)
with open(whitelist_path, "r") as f:
self.whitelist = np.load(f)
with tf.variable_scope("jpeg_distort"):
self.input, image = image_util.add_jpeg_decoding(img_stats.depth)
resized = image_util.resize(image, img_stats.size, img_stats.size)
norm_l = image_util.normalize(image_util.brighten(resized, 5),
img_stats.mean, img_stats.std)
norm_r = image_util.normalize(image_util.brighten(resized, 5),
img_stats.mean, img_stats.std)
self.l_output = tf.expand_dims(tf.squeeze(norm_l), 0)
self.r_output = tf.expand_dims(
tf.image.flip_left_right(tf.squeeze(norm_r)), 0)
self.image_whitelist = np.unique(self.whitelist)
def image_to_array(img, size):
#show_image(img)
side = img.shape[:1]
side = side[0] / 9
result = [[0 for i in range(9)] for j in range(9)]
for i in range(9):
for j in range(9):
tl = (i * side, j * side) # Top left corner
br = ((i + 1) * side, (j + 1) * side) # Bottom right corner
digit = crop_digit(img, (tl, br), size)
if (digit is not None):
digit = resize(digit, size, 4)
#show_image(digit)
feature = image_to_feature(digit)
if len(feature) != 0:
predicted = model.predict([feature])[0]
#print(predicted)
result[j][i] = int(predicted)
else:
result[j][i] = int(0)
return result
def create_mosaic_lines(files_rgb: List[FileInfo], img_target, thumb_width,
thumb_height, thumb_width_count, thumb_height_count,
img_part_width, img_part_height):
lines = []
for i in range(0, thumb_height_count - 1):
line = None
for j in range(0, thumb_width_count - 1):
x = j * img_part_width
y = i * img_part_height
rect_rgb = get_int_from_rgb(
rect_median_color(img_target, x, y, img_part_width,
img_part_height))
rect_best_fit = find_best_fit(rect_rgb, files_rgb)
resized = resize(cv2.imread(rect_best_fit), thumb_width,
thumb_height)
if line is None:
line = resized
else:
line = np.concatenate((line, resized), axis=1)
lines.append(line)
return lines
def add_resizer(self, ratio):
resizer = lambda im: image_util.resize(im, ratio)
self.ops.append(resizer)
def convert(self, img_path, objs):
"""
Input:
img_path: 画像ファイルのパス
objs: オブジェクトbounding boxの情報を含んだ配列。各オブジェクトは(xmin, ymin, xmax, ymax, class_id)
Output:
(3, target_size, target_size)の画像, bboxの配列
bboxの配列は長さ n_grid**2 で、各要素は
(中心点のgrid内x, 中心点のgrid内y, 幅, 高さ, class_id)
"""
orig_img = Image.open(img_path)
w, h = orig_img.size
if self.augment:
if random() > 0.5:
t_img, t_boxes = orig_img, objs
else:
t_img, t_boxes = horizontal_flip(orig_img, objs)
max_noise = min(w, h) * 0.2
w_noise = int(max_noise * (random() - 0.5))
h_noise = int(max_noise * (random() - 0.5))
t_img, t_boxes = resize(t_img,
t_boxes,
self.target_size,
w_noise,
h_noise,
min_bbox_pixel=3)
np_img = self.color_noise(
numpy.asarray(t_img.convert('RGB'),
dtype=numpy.float32)).transpose(2, 0, 1)
else:
t_img, t_boxes = resize(orig_img,
objs,
self.target_size,
min_bbox_pixel=3)
np_img = numpy.asarray(t_img.convert('RGB'),
dtype=numpy.float32).transpose(2, 0, 1)
np_boxes = numpy.zeros((self.n_grid**2, 5 + self.n_classes),
dtype=numpy.float32)
np_boxes[:, 5:] = self.class_weights
for box in t_boxes:
try:
w_min, h_min, w_max, h_max, cls_id = box
grid_idx, x_grid_index, y_grid_index, grid_center_x, grid_center_y, relative_w, relative_h = self.grid_retriever.grid_position(
w_min, h_min, w_max, h_max)
np_boxes[grid_idx][0] = grid_center_x
np_boxes[grid_idx][1] = grid_center_y
np_boxes[grid_idx][2] = relative_w
np_boxes[grid_idx][3] = relative_h
np_boxes[grid_idx][4] = cls_id
if not self.rescore_neighbour:
continue
neighbours = []
if grid_center_x < 0.5 and x_grid_index > 0:
neighbours.append((grid_idx - 1, grid_center_x / 0.5))
if grid_center_y < 0.5 and y_grid_index > 0:
_distance = (grid_center_x**2 + grid_center_y**2)**0.5
neighbours.append((grid_idx - self.n_grid - 1,
min(_distance / 0.5, 1.0)))
elif grid_center_y > 0.5 and y_grid_index < self.n_grid - 1:
_distance = (grid_center_x**2 +
(1.0 - grid_center_y)**2)**0.5
neighbours.append((grid_idx + self.n_grid - 1,
min(_distance / 0.5, 1.0)))
elif grid_center_x > 0.5 and x_grid_index < self.n_grid - 1:
neighbours.append(
(grid_idx + 1, (1.0 - grid_center_x) / 0.5))
if grid_center_y < 0.5 and y_grid_index > 0:
_distance = ((1.0 - grid_center_x)**2 +
grid_center_y**2)**0.5
neighbours.append((grid_idx - self.n_grid + 1,
min(_distance / 0.5, 1.0)))
elif grid_center_y > 0.5 and y_grid_index < self.n_grid - 1:
_distance = ((1.0 - grid_center_x)**2 +
(1.0 - grid_center_y)**2)**0.5
neighbours.append((grid_idx + self.n_grid + 1,
min(_distance / 0.5, 1.0)))
if grid_center_y < 0.5 and y_grid_index > 0:
neighbours.append(
(grid_idx - self.n_grid, grid_center_y / 0.5))
elif grid_center_y > 0.5 and y_grid_index < self.n_grid - 1:
neighbours.append(
(grid_idx + self.n_grid, (1.0 - grid_center_y) / 0.5))
for grid_i, w in neighbours:
np_boxes[grid_i,
5 + cls_id] = min(np_boxes[grid_i, 5 + cls_id], w)
except Exception as e:
print(e)
print(box)
print(
self.grid_retriever.grid_position(w_min, h_min, w_max,
h_max))
return np_img, np_boxes