def process_first_line(img, is_country):
img_h, img_w, _ = img.shape
kernel = np.ones((25, 25), np.uint8)
thresh = get_threshold_img(img, kernel)
contour_boxes = get_contour_boxes(thresh)
avg = statistics.mean(map(lambda t: t[-1] * t[-2], contour_boxes))
boxes_copy = copy.deepcopy(contour_boxes)
for box in boxes_copy:
if box[-1] * box[-2] < avg / 3:
contour_boxes.remove(box)
contour_boxes.sort(key=lambda t: t[0])
list_distance = []
for index, box in enumerate(contour_boxes):
current_x = box[0] + box[2]
if index < len(contour_boxes) - 1:
next_x = contour_boxes[index + 1][0]
list_distance.append(next_x - current_x)
avg = statistics.mean(list_distance)
list_copy = copy.deepcopy(list_distance)
list_copy.sort(reverse=True)
if len(list_copy) > 1 and list_copy[0] > 3 * list_copy[1]:
max_index = list_distance.index(list_copy[0])
contour_boxes = contour_boxes[max_index + 1:]
x, y, w, h = find_max_box(contour_boxes)
img = img[0:img_h, x:img_w]
return get_text(img)
else:
return strip_label_and_get_text(img, is_country)
def get_text_from_two_lines(img, box):
x0, y0, x1, y1 = box
img = img[y0:y1, x0:x1]
kernel = np.ones((25, 25), np.uint8)
thresh = get_threshold_img(img, kernel)
height, width = thresh.shape
kernel = cv2.getStructuringElement(cv2.MORPH_RECT, (5, 5))
dilation = cv2.dilate(thresh, kernel, iterations=1)
contour_boxes = get_contour_boxes(dilation)
avg = statistics.mean(map(lambda t: t[-1] * t[-2], contour_boxes))
boxes_copy = copy.deepcopy(contour_boxes)
for box in boxes_copy:
box_height = box[1] + box[3]
height_lim = 0.9 * height
if box[1] > height_lim:
contour_boxes.remove(box)
elif box_height == height and box[1] > 0.8 * height:
contour_boxes.remove(box)
elif box[-1] * box[-2] < avg / 3:
contour_boxes.remove(box)
x, y, w, h = find_max_box(contour_boxes)
if h < 55:
return (x0 + x, y0 + y, x0 + x + w + 5, y0 + y + h + 5)
else:
crop_img = thresh[y:y + h, x:width]
height, width = crop_img.shape
hist = cv2.reduce(crop_img, 1, cv2.REDUCE_AVG).reshape(-1)
hist = uppers = [hist[y] for y in range(height // 3, 2 * height // 3)]
line = uppers.index(min(uppers)) + height // 3
first_line = (x0 + x, y0 + y, x0 + x + w, y0 + y + line)
second_line = (x0 + x, y0 + y + line, x0 + x + w, y0 + y + h)
return [first_line, second_line]
def get_id_numbers_text(img):
height_img, width_img, _ = img.shape
if height_img < 20:
img = cv2.resize(img, (2 * width_img, 2 * height_img),
interpolation=cv2.INTER_CUBIC)
height, width, _ = img.shape
kernel = np.ones((height // 2, height // 2), np.uint8)
thresh = get_threshold_img(img, kernel)
boxes = get_contour_boxes(thresh)
boxes_copy = copy.deepcopy(boxes)
for box in boxes_copy:
if box[3] < 0.4 * height:
boxes.remove(box)
boxes.sort(key=lambda t: t[0])
list_number = []
for box in boxes:
x, y, w, h = box
if x < 2 or x + w + 2 > width:
continue
thresh_number = thresh[0:height, x - 2:x + w + 2]
normal_number = img[0:height, x - 2:x + w + 2]
list_number.append((thresh_number, normal_number))
numbers = gather_results(
[run_item(get_each_number, item) for item in list_number])
text = ''.join(numbers)
return text[-12:]
def detect_info(img):
img, face = cropout_unimportant_part(img)
orig = img.copy()
img, ratio = resize_img_by_height(img)
label_img = crop_label(img)
kernel = cv2.getStructuringElement(cv2.MORPH_RECT, (15, 15))
threshold_img = get_threshold_img(label_img, kernel)
kernel = cv2.getStructuringElement(cv2.MORPH_RECT,
(label_img.shape[1] // 2, 5))
dilation = cv2.dilate(threshold_img, kernel, iterations=1)
contour_boxes = get_contour_boxes(dilation)
contour_boxes.sort(key=lambda t: t[2] * t[3], reverse=True)
contour_boxes = contour_boxes[:5]
info_list = get_info_list(img, contour_boxes)
# get number part
x, y, _, _ = info_list[0]
number_box = (0, 0, img.shape[1], info_list[0][1])
number_box = get_main_text(img, number_box, 5)
number_img = get_img_from_box(orig, ratio, number_box)
# get name part
name_box = info_list[0]
name_box = get_name(img, get_main_text(img, name_box, 5))
name_img = get_img_from_box(orig, ratio, name_box, padding=2)
name_img = cut_blank_part(name_img)
# get dob part
dob_box = info_list[1]
dob_box = get_main_text(img, dob_box, 5)
dob_img = get_img_from_box(orig, ratio, dob_box)
# get gender_and national part
gender_and_nationality_box = info_list[2]
gender_and_nationality_box = get_main_text(img, gender_and_nationality_box,
5)
gender_n_nation_img = get_img_from_box(orig,
ratio,
gender_and_nationality_box,
padding=2)
h, w, _ = gender_n_nation_img.shape
gender_img = gender_n_nation_img[0:h, 0:int(w / 3)]
nation_img = gender_n_nation_img[0:h, int(w / 3):int(w)]
nation_img = cut_blank_part(nation_img)
# get country part
country_box = info_list[3]
x, y, x1, y1 = country_box
last_y = gender_and_nationality_box[-1]
country_img = process_result(orig, ratio,
get_two_lines_img(img,
(x, last_y, x1, y1)))[0]
country_result = get_text_from_two_lines(img, (x, last_y, x1, y1))
country_img_list = process_result(orig, ratio, country_result)
address_box = info_list[4]
x, y, x1, y1 = address_box
last_y = get_last_y(country_result)
address_img = process_result(orig, ratio,
get_two_lines_img(img,
(x, last_y, x1, y1)))[0]
result = get_text_from_two_lines(img, (x, last_y, x1, y1))
address_img_list = process_result(orig, ratio, result)
return face, number_img, name_img, dob_img, gender_img, nation_img, country_img, \
address_img, country_img_list, address_img_list
def get_name(img, box):
x0, y0, x1, y1 = box
img = img[y0:y1, x0:x1]
height, width, _ = img.shape
kernel = cv2.getStructuringElement(cv2.MORPH_RECT, (25, 25))
thresh_img = get_threshold_img(img, kernel)
contour_boxes = get_contour_boxes(thresh_img)
contour_boxes = remove_smaller_area(contour_boxes, width)
contour_boxes = remove_name_label(contour_boxes, width)
contour_boxes.sort(key=lambda t: t[0])
x, y, w, h = find_max_box(contour_boxes)
return (x0 + x, y0 + y, x0 + x + w, y0 + y + h)
def get_main_text(img, box, kernel_height):
x0, y0, x1, y1 = box
img = img[y0:y1, x0:x1]
kernel = cv2.getStructuringElement(cv2.MORPH_RECT, (25, 25))
thresh = get_threshold_img(img, kernel)
kernel = cv2.getStructuringElement(cv2.MORPH_RECT,
(thresh.shape[1], kernel_height))
dilation = cv2.dilate(thresh, kernel, iterations=1)
contour_boxes = get_contour_boxes(dilation)
max_box = max(contour_boxes, key=lambda tup: tup[2] * tup[3])
x, y, w, h = max_box
return (x0 + x, y0 + y, x0 + x + w, y0 + y + h)
def get_information_x_axis(img):
img, ratio = resize_img_by_height(img)
h, w, _ = img.shape
img_resize = img[100:400, int(0.25 * w):int(0.4 * w)]
kernel = cv2.getStructuringElement(cv2.MORPH_RECT, (100, 100))
thresh = get_threshold_img(img_resize, kernel)
kernel = cv2.getStructuringElement(cv2.MORPH_RECT, (5, h))
dilation = cv2.dilate(thresh, kernel, iterations=1)
cnts = get_contour_boxes(dilation)
cnts_copy = copy.deepcopy(cnts)
for cnt in cnts_copy:
if cnt[0] < 0.1 * img_resize.shape[1]:
cnts.remove(cnt)
max_cnt = max(cnts, key=lambda x: x[-1] * x[-2])
return int((max_cnt[0] - 5 + 0.25 * w) * ratio)
def cut_blank_part(img, padding=5):
img_h, img_w, _ = img.shape
kernel = np.ones((25, 25), np.uint8)
thresh = get_threshold_img(img, kernel)
contour_boxes = get_contour_boxes(thresh)
avg = statistics.mean(map(lambda t: t[-1], contour_boxes))
boxes_copy = copy.deepcopy(contour_boxes)
for box in boxes_copy:
if box[-1] < avg / 2:
contour_boxes.remove(box)
elif box[1] > img_h / 2 and box[0] < img_w / 10:
contour_boxes.remove(box)
elif box[1] < img_h / 10 and box[-1] < img_h / 5:
contour_boxes.remove(box)
x, y, w, h = find_max_box(contour_boxes)
new_width = x + w + padding
if new_width > img_w:
new_width = img_w
return img[0:img_h, x:new_width]
def get_information_y_axis(img):
img, ratio = resize_img_by_width(img)
h, w, _ = img.shape
img_resize = img[0:int(0.4 * h), 125:w]
gray = cv2.cvtColor(img_resize, cv2.COLOR_BGR2GRAY)
kernel = np.ones((25, 25), np.uint8)
blackhat = cv2.morphologyEx(gray, cv2.MORPH_BLACKHAT, kernel)
thresh = cv2.threshold(blackhat, 0, 255,
cv2.THRESH_BINARY + cv2.THRESH_OTSU)[1]
kernel = cv2.getStructuringElement(cv2.MORPH_RECT, (20, 3))
dilation = cv2.dilate(thresh, kernel, iterations=1)
cnts = get_contour_boxes(dilation)
cnts_copy = copy.deepcopy(cnts)
for cnt in cnts_copy:
if cnt[1] + cnt[-1] > 0.95 * img_resize.shape[0]:
cnts.remove(cnt)
elif cnt[-2] < 150:
cnts.remove(cnt)
max_cnt = max(cnts, key=lambda x: x[1])
return int((max_cnt[1] - 5) * ratio)
def get_two_lines_img(img, box):
x0, y0, x1, y1 = box
img = img[y0:y1, x0:x1]
kernel = np.ones((25, 25), np.uint8)
thresh = get_threshold_img(img, kernel)
height, width = thresh.shape
kernel = cv2.getStructuringElement(cv2.MORPH_RECT, (5, 5))
dilation = cv2.dilate(thresh, kernel, iterations=1)
contour_boxes = get_contour_boxes(dilation)
avg = statistics.mean(map(lambda t: t[-1] * t[-2], contour_boxes))
boxes_copy = copy.deepcopy(contour_boxes)
for box in boxes_copy:
box_height = box[1] + box[3]
height_lim = 0.9 * height
if box[1] > height_lim:
contour_boxes.remove(box)
elif box_height == height and box[1] > 0.8 * height:
contour_boxes.remove(box)
elif box[-1] * box[-2] < avg / 3:
contour_boxes.remove(box)
x, y, w, h = find_max_box(contour_boxes)
return (x0 + x, y0 + y, x0 + x + w + 5, y0 + y + h + 5)
