def check_down(self, path):
colors_check = np.zeros((15, 6), dtype=np.int)
points_check = np.zeros((15, 6), dtype=np.int)
img = cv2.imread(path)
crop = img[400:1500, 1300:1900]
# crop = cv2.flip(crop,-1)
saver(crop, "C")
crop_gray = cv2.cvtColor(crop, cv2.COLOR_BGR2GRAY)
circles = cv2.HoughCircles(crop_gray,
cv2.HOUGH_GRADIENT,
1,
20,
param1=50,
param2=30,
minRadius=15,
maxRadius=34)
draw = self._drawCircles(crop, circles)
saver(draw, "D_%s" % path[-8:-4])
one = circles[0]
one = sorted(one, key=cmp_to_key(_sortSmall))
radius = 30
count = 0
points = 0
for c in one:
# print(c)
c = np.array(c).astype('int')
circle = crop[c[1] - radius:c[1] + radius,
c[0] - radius:c[0] + radius, :]
up_circle = circle[0:radius, radius:2 * radius]
down_circle = circle[radius:2 * radius, 0:radius]
# saver(circle,"C_%d"%count)
# saver(down_circle, "D_%d"%count)
# saver(up_circle, "U_%d"%count)
color = checkAllColor(circle)
points = self._checkPont(up_circle, 1) + self._checkPont(
down_circle, 0)
colors_check[count // 6][count % 6] = color
points_check[count // 6][count % 6] = points
count += 1
for c in range(count, 90):
colors_check[count // 6][count % 6] = -1
points_check[count // 6][count % 6] = -1
count += 1
return colors_check, points_check
def check_card(self, path):
colors_check = np.zeros((15, 6), dtype=np.int)
img = cv2.imread(path)
crop = img[195:1215, 765:1185]
crop_gray = cv2.cvtColor(crop, cv2.COLOR_BGR2GRAY)
crop_gray = cv2.GaussianBlur(crop_gray, (5, 5), 0)
circles = cv2.HoughCircles(crop_gray,
cv2.HOUGH_GRADIENT,
1,
40,
param1=45.5,
param2=20,
minRadius=15,
maxRadius=35)
draw = self._drawCircles(crop, circles)
saver(draw, "D_%s" % path[-8:-4])
one = circles[0]
one = sorted(one, key=cmp_to_key(_sortSmall))
radius = 30
count = 0
points = 0
for c in one:
# print(c)
try:
c = np.array(c).astype('int')
circle = crop[c[1] - radius:c[1] + radius,
c[0] - radius:c[0] + radius, :]
color = checkAllColor(circle)
colors_check[count // 6][count % 6] = color
count += 1
except Exception as e:
print("run out of boundry")
break
for c in range(count, 90):
colors_check[count // 6][count % 6] = -1
count += 1
return colors_check
else:
return 2
def checkAllColor(img):
blue_rate = checkBlue(img).mean()
red_rate = checkRed(img).mean()
green_rate = checkGreen(img).mean()
if red_rate < 30 and blue_rate < 30 and green_rate<30:
return -1
if blue_rate>red_rate and blue_rate>green_rate:
return 2
if red_rate>blue_rate and red_rate>green_rate:
return 1
if green_rate>blue_rate and green_rate>red_rate:
return 3
return -1
if __name__ == "__main__":
path = "outer/test_C_22.png"
img = cv2.imread(path)
r = checkGreen(img)
print("Green: ", r.mean())
saver(r,"GREEN")
print("Blue: ", checkBlue(img).mean())
saver(checkBlue(img),"Blue")
print("Green", checkGreen(img).mean())
saver(checkRed(img),"Red")
if args.use_metric:
model_metric = ArcFaceLoss(in_features=int(args.layers[-1]),
out_features=sc_dataset_train.num_class)
model_metric.to(device)
parameters.append({'params': model_metric.parameters()})
model_name = 'MLP_{}_metric'.format("-".join(args.layers))
else:
model_metric = None
model_name = 'MLP_{}'.format("-".join(args.layers))
optimizer = torch.optim.SGD(parameters,
lr=args.learning_rate,
momentum=args.sgd_momentum)
embedding_writer = SummaryWriter(results_path)
saver = saver(save_path=results_path,\
model_name=model_name,\
writer=embedding_writer,\
args=args)
results = train(dataloaders=dataloader_dict,\
model=model,\
model_metric=model_metric,\
paired=True,
criterion=criterion,\
optimizer=optimizer,\
device=device,\
saver=saver,\
num_epoch=args.num_epoch)
kernel = np.ones((4, 4), np.uint8)
def clean_noise(im, type):
if type == 1:
green_mask = checkGreen(im)
x2, y2 = green_mask.shape
mix_mask = cv2.rectangle(im.copy(), (0, 0), (y2, x2), (255, 255, 255),
-1)
green_mask = cv2.erode(green_mask, kernel, iterations=1)
green_mask = cv2.dilate(green_mask, kernel, iterations=1)
_, green_mask_I = cv2.threshold(green_mask.copy(), 10, 255,
cv2.THRESH_BINARY_INV)
_, green_mask = cv2.threshold(green_mask.copy(), 10, 255,
cv2.THRESH_BINARY)
clean = cv2.bitwise_and(im, im.copy(), mask=green_mask_I)
mix_mask = cv2.bitwise_and(mix_mask, mix_mask.copy(), mask=green_mask)
clean = cv2.add(clean, mix_mask)
return clean
return im
if __name__ == "__main__":
path = './outer/test_CROP.png'
im = cv2.imread(path)
cleaned = clean_noise(im, 1)
saver(cleaned, "G")
label,
num_batch_val * (epoch - 1) + batch,
dataformats='NHWC')
writer_val.add_image('image',
image,
num_batch_val * (epoch - 1) + batch,
dataformats='NHWC')
writer_val.add_image('output',
output,
num_batch_val * (epoch - 1) + batch,
dataformats='NHWC')
# Tensorboard 저장하기 (손실함수 저장)
writer_val.add_scalar('loss', np.mean(loss_arr), epoch)
if epoch // 5 == 0:
saver(ckpt_dir=ckpt_dir, net=net, optim=optim, epoch=epoch)
writer_train.close()
writer_val.close()
## 데이터 시각화
data = dataset_train.__getitem__(0)
image = data['image']
label = data['label']
plt.subplot(121)
plt.imshow(image.squeeze()) # squeeze를 하지 않으면 데이터의 shape가 (512, 512, 1)이 된다.
# squeeze로 차원 하나를 제거하여 (512,512)로 만든다.
plt.subplot(122)
plt.imshow(label.squeeze())
def check_up(self, path):
result = np.zeros((self.size[1]), dtype=np.int)
err = 0
y_index = 0 #第几行
count = 0 #第几个
ll = 0
outer_side = 0
# Read
img = cv2.imread(path)
crop = img[:, self.range[2]:self.range[3], :]
if self.type not in [5]:
crop = cv2.flip(crop, -1)
crop_gray = cv2.cvtColor(crop, cv2.COLOR_BGR2GRAY)
circles = self._findCircles(crop_gray)
# Logger
draw = self._drawCircles(crop, circles)
saver(draw, "D_%s" % path[-8:-4])
if path[-8:-4] in ["110", "184"]:
err = 1
one = circles[0]
one = sorted(one, key=cmp_to_key(_sortCircle))
y_min = one[0][1]
x_max = one[0][0]
# 过滤Type1的两个噪声
if self.type == 1:
one = one[2:]
for c in one:
count += 1
c = np.array(c).astype('int')
circle = crop[c[1] - self.radius:c[1] + self.radius,
c[0] - self.radius:c[0] + self.radius, :]
# 原则上噪声与可识别区域间隔三行
if c[1] - y_min > self.radius * 6:
break
color = checkColor(circle)
if _abs(c[1] - y_min) > self.radius * 2 - 20:
y_index += 1
x_max = c[0] #换行了，代表最右边的坐标x—max更新
y_min = c[1]
input_index = y_index #如果没有问题,用y_index作为结果添加的位置
#如果这个圆圈是在左边的圆圈，并且右边没有圆圈(就是说这个圆圈的x等于代表第一个圆圈的x_max)
if outer_side != 0 and c[0] == x_max and abs(
c[0] - outer_side) < self.radius:
input_index = ll
# 这个圆圈和它右边的链接不上，应该是换行来的圆圈
if x_max - c[0] > self.radius * 3:
input_index = ll
else:
x_max = c[0]
add = 0
if color == 1:
add = 1
else:
add = -1
#如果两个圆圈颜色不同，就是换行来的
if abs(result[y_index]) == 5 and abs(result[y_index] + add) < abs(
result[y_index]) + abs(add):
input_index = ll
if y_index - input_index >= 2 and abs(
result[input_index + 1]) >= 5 and abs(
result[input_index]) >= 7:
print("GG")
if abs(add + result[input_index + 1]) == abs(add) + abs(
result[input_index + 1]):
input_index += 1
#进行叠加计算
result[input_index] += add
if abs(result[y_index]) >= self.size[0]:
ll = y_index
outer_side = c[0]
#识别报错的东西
for i in range(0, y_index):
if abs(result[i]) >= 6:
for j in range(i + 1, y_index):
#print("I: ", i, "J: ",j, result[i], result[j])
if abs(result[i] + result[j]) >= abs(result[i]) + abs(
result[j]) and abs(result[j]) >= 6:
if abs(result[i]) - 5 >= j - i:
err = 1
break
return result, one, err