def ha_SignCompare(image, text):
img = Image.open(image).convert('RGB')
threshold = 200
limg = img.convert('L').point(lambda p: 0 if p > threshold else 1, '1')
img_data = np.array(limg, dtype=np.int32)
dire_sum_0 = img_data.sum(axis=0)
dire_sum_1 = img_data.sum(axis=1)
no0_dir0 = np.where(dire_sum_0>0)[0]
no0_dir1 = np.where(dire_sum_1>0)[0]
direction = 1
if no0_dir0[-1] - no0_dir0[0] < no0_dir1[-1] - no0_dir1[0]:
direction = 0
code, message, char_count, char_list = 0, '', 0, []
if direction == 0:
borders = find_best_borders_projection(img, text, direction = direction, min_sum=1, min_gap=1)
if borders:
for border in borders:
pred_char, conf, toplist = image_prediction(img.crop(border), topk = 30)
char_list.append(toplist)
else:
borders = find_best_borders_connect(img, text, direction = direction)
if borders:
for border in borders:
pred_char, conf, toplist = image_prediction(img.crop(border), topk = 30)
char_list.append(toplist)
else:
borders = find_best_borders_gas(img, text, direction = direction)
if borders:
for border in borders:
pred_char, conf, toplist = image_prediction(img.crop(border), topk = 30)
char_list.append(toplist)
else:
borders = find_best_borders_gas(img, text, direction = direction)
if borders:
for border in borders:
pred_char, conf, toplist = image_prediction(img.crop(border), topk = 30)
char_list.append(toplist)
# show_borders(img, borders)
if char_list:
return 0, 'Successful', len(char_list), char_list
else:
return -1, 'Failed', 0, []
def ha_SignCompare(image, text):
img = Image.open(image).convert('RGB')
direction = 0 if img.size[0] < img.size[1] else 1
name_num = len(text)
code, message, char_count, char_list = 0, '', 0, []
if direction == 0:
borders = find_best_borders_projection(img,
text,
direction=direction,
min_sum=1,
min_gap=1)
if borders:
for border in borders:
pred_char, conf, toplist = image_prediction(img.crop(border),
topk=30)
char_list.append(toplist)
else:
borders = find_best_borders_connect(img, text, direction=direction)
if borders:
for border in borders:
pred_char, conf, toplist = image_prediction(
img.crop(border), topk=30)
char_list.append(toplist)
else:
borders = find_best_borders_gas(img, text, direction=direction)
if borders:
for border in borders:
pred_char, conf, toplist = image_prediction(
img.crop(border), topk=30)
char_list.append(toplist)
else:
borders = find_best_borders_gas(img, text, direction=direction)
if borders:
for border in borders:
pred_char, conf, toplist = image_prediction(img.crop(border),
topk=30)
char_list.append(toplist)
if char_list:
return 0, 'Successful', len(char_list), char_list
else:
return -1, 'Failed', 0, []
def find_best_borders_connect(img, name, threshold = 200, direction = 0, overlap = 0.85, net_type = 'S'):
limg = img.convert('L').point(lambda p: 0 if p > threshold else 1, '1')
img_data = np.array(limg, dtype=np.int32)
direction_sum = img_data.sum(axis=direction)
up, down = np.where(direction_sum>0)[0][0], np.where(direction_sum>0)[0][-1]
verdire_sum = img_data.sum(axis=1-direction)
length = np.where(verdire_sum>0)[0][-1] - np.where(verdire_sum>0)[0][0]
merge_area = find_connect(img, direction = direction, overlap = overlap)
area_num = len(merge_area)
char_num = len(name)
name_list = list(name)
char_min = length / char_num / 3
char_max = min(length / char_num * 2, length / 3 * 2)
# 连笔严重，无法切割出合适的连通区域
if char_num > area_num:
return []
elif char_num == area_num:
for area in merge_area:
if (area[1] - area[0] <= char_min or area[1] - area[0] >= char_max ):
return []
max_seg = [-1] + [i for i in range(char_num)]
else:
for area in merge_area:
if area[1] - area[0] >= char_max:
return []
confM = [[0 for i in range(area_num)] for j in range(area_num)]
charM = [[' ' for i in range(area_num)] for j in range(area_num)]
charConfM = [[[0] * char_num for i in range(area_num)] for j in range(area_num)]
for i in range(area_num):
for j in range(i, area_num):
if (merge_area[j][1] - merge_area[i][0] >= char_min and
merge_area[j][1] - merge_area[i][0] <= char_max ):
if direction == 0:
border = [up, merge_area[i][0], down, merge_area[j][1]]
elif direction == 1:
border = [merge_area[i][0], up, merge_area[j][1], down]
pred_img = img.crop(border)
# display(pred_img)
charM[i][j], confM[i][j], _, charConfM[i][j] = image_prediction(pred_img, topk = 1, labels = name_list, net_type = net_type)
_, max_conf, max_seg = find_seg_point(area_num, char_num, confM, charConfM, 2, 0, char_num, [-1], 0, [-1])
if len(max_seg) != char_num + 1:
return []
borders = []
for i in range(char_num):
if direction == 0:
borders.append([up, merge_area[max_seg[i]+1][0], down, merge_area[max_seg[i+1]][1]])
elif direction == 1:
borders.append([merge_area[max_seg[i]+1][0], up, merge_area[max_seg[i+1]][1], down])
return borders
def find_best_borders_gas(img, name, threshold = 200, direction = 0, gas_std = 8, net_type = 'S'):
name_list = list(name)
char_num = len(name)
limg = img.convert('L').point(lambda p: 0 if p > threshold else 1, '1')
img_data = np.array(limg, dtype=np.int32)
direction_sum = img_data.sum(axis=direction)
up, down = np.where(direction_sum>0)[0][0], np.where(direction_sum>0)[0][-1]
verdire_sum = img_data.sum(axis=1-direction)
gas8 = gaussian_filter1d(verdire_sum, gas_std)
local_min = find_local_min(gas8)
length = local_min[-1] - local_min[0]
min_num = len(local_min)
char_min = length / char_num / 2
char_max = min(length / char_num * 2, length / 3 * 2)
count = 0
if min_num-1 < char_num:
return []
elif min_num-1 == char_num:
for i in range(1, min_num):
if local_min[i] - local_min[i-1] <= char_min or local_min[1] - local_min[0] >= char_max:
return []
max_seg = [i for i in range(min_num)]
else:
confM3 = [[0 for i in range(min_num)] for j in range(min_num)]
charM3 = [[' ' for i in range(min_num)] for j in range(min_num)]
charConfM3 = [[[0] * char_num for i in range(min_num)] for j in range(min_num)]
for i in range(min_num):
for j in range(i+1, min_num):
if(local_min[j] - local_min[i] >= char_min and local_min[j] - local_min[i] <= char_max):
if direction == 0:
border = [up, local_min[i], down, local_min[j]]
elif direction == 1:
border = [local_min[i], up, local_min[j], down]
pred_img = img.crop(border)
# display(pred_img)
count += 1
charM3[i][j], confM3[i][j], toplist, charConfM3[i][j] = image_prediction(pred_img, topk = 1, labels=name_list, net_type = net_type)
_, max_conf, max_seg = find_seg_point(min_num, char_num, confM3, charConfM3, 3, 0, char_num, [0], 0, [0])
if len(max_seg) != char_num + 1:
return []
# print(name, ", minpoint count: ", min_num, ", predict count: ", count)
borders = []
for i in range(char_num):
if direction == 0:
borders.append([up, local_min[max_seg[i]], down, local_min[max_seg[i+1]]])
elif direction == 1:
borders.append([local_min[max_seg[i]], up, local_min[max_seg[i+1]], down])
return borders