def prob_map(model_path, img_ori):
model = torch.jit.load(model_path)
model = torch.nn.DataParallel(model).cuda()
with torch.no_grad():
model.eval()
try:
img, _ = cv2.decolor(img_ori)
img = np.array(img, dtype=np.float32)
img = (img - 128.0) / 128.0
img = torch.from_numpy(img.copy())
print("img:", img.size())
img = img.unsqueeze(0)
input = img.unsqueeze(0)
input = input.cuda()
output = model(input)
img_size_1, img_size_2 = input.size(2), input.size(3)
output = output.squeeze()
output = output.cpu().numpy()
mask = np.zeros((img_size_1, img_size_2))
x, y = np.where(output > 0.5)
mask[x, y] = 255
print('end')
except:
img, _ = cv2.decolor(img_ori)
w, h = img.shape
img = np.array(img, dtype=np.float32)
img = (img - 128.0) / 128.0
img = torch.from_numpy(img.copy())
img = img.unsqueeze(0)
img = img.unsqueeze(0)
print("img:", img.size())
img_slice1 = img[:, :, 0:w // 2, 0:h // 2]
img_slice2 = img[:, :, w // 2:, 0:h // 2]
img_slice3 = img[:, :, 0:w // 2, h // 2:]
img_slice4 = img[:, :, w // 2:, h // 2:]
# print("img slice1:", img_slice1.size())
input1 = img_slice1.cuda()
output1 = model(input1)
input2 = img_slice2.cuda()
output2 = model(input2)
input3 = img_slice3.cuda()
output3 = model(input3)
input4 = img_slice4.cuda()
output4 = model(input4)
output = torch.cat((torch.cat(
(output1, output2), 2), torch.cat((output3, output4), 2)), 3)
output = output.squeeze()
output = output.cpu().numpy()
mask = np.zeros((w, h))
x, y = np.where(output > 0.5)
mask[x, y] = 255
print('end oversize')
return mask
def decolor(img):
gray = np.zeros(img.shape, np.uint8)
gray = cv2.cvtColor(gray, cv2.CV_8UC1)
color_boost = np.zeros(img.shape, np.uint8)
color_boost = cv2.cvtColor(color_boost, cv2.CV_8UC3)
decolor = cv2.decolor(img, gray, color_boost)
return decolor
def demo(device_num, ext='jpg', delay=1):
cap = setup_webcam(device_num)
# cascade = cv2.CascadeClassifier('cascades/haarcascade_frontalface_alt2.xml')
cascade = cv2.CascadeClassifier(
'cascades/haarcascade_frontalface_default.xml')
p = Predictor(quant=True)
inference_result = None
while True:
ret, frame = cap.read()
key = cv2.waitKey(delay) & 0xFF
if key == ord('q'):
break
# Execute Face Detection
faces = detect_face(frame, cascade)
# Draw Detected Face
for (x, y, w, h) in faces:
# Inference test data
face_img = frame[y:y + h, x:x + w]
face_img, _ = cv2.decolor(face_img)
face_img = cv2.resize(face_img,
dsize=(dataset.IMAGE_SIZE,
dataset.IMAGE_SIZE))
# cv2.imshow(WINDOW_NAME, face_img)
start = time.time()
inference_result = p(np.array(face_img, dtype=np.uint8))
end = time.time()
measured_time = (end - start) * 1000
frame = cv2.rectangle(frame, (x, y), (x + w, y + h), (255, 0, 0),
2)
if inference_result is not None:
expression = dataset.CLASS_NAME[np.argmax(inference_result)]
if not np.argmax(inference_result) in [0, 3, 5, 6]:
expression = "None"
percentage = ": {0:.2f}".format(
inference_result[np.argmax(inference_result)] * 100)
cv2.putText(frame, expression + ": " + percentage, (10, 400),
cv2.FONT_HERSHEY_PLAIN, 2, (0, 0, 255), 2, cv2.LINE_AA)
cv2.putText(
frame,
"{0:.2f} ms, {1:.2f} fps".format(measured_time,
1000.0 / measured_time),
(10, 440), cv2.FONT_HERSHEY_PLAIN, 2, (0, 255, 0), 2,
cv2.LINE_AA)
cv2.imshow(WINDOW_NAME, frame)
cv2.destroyWindow(WINDOW_NAME)
def threshold_otu(image: np.ndarray) -> np.ndarray:
"""大津の2値化
Args:
image: 変換元の画像
Returns:
変換後の画像
"""
#: 退色処理
image_gray, _ = cv2.decolor(image)
#: 2値化
image_gray[image_gray < 128] = 16
image_gray[image_gray >= 128] = 240
return image_gray
def gray(img):
g, _ = cv2.decolor(img)
return g
import cv2
import numpy as np
image = None
dec = None
image = cv2.imread('media/lena.jpg',1)
dec = (cv2.decolor(image))
cv2.imshow('gray',dec[0])
cv2.imshow('color',dec[1])
if cv2.waitKey(0)&0xff == 27:
pass
cv2.destroyAllWindows()
def monochrome(self): copy_img = self.img.copy() # グレースケールはやり方がいくつかある。 # self.reimg = cv2.cvtColor(copy_img, cv2.COLOR_BGR2GRAY) # RGB2〜 でなく BGR2〜 を指定 self.reimg, _ = cv2.decolor(copy_img)
def __getitem__(self, idx):
mask = Image.open(self.mask[idx])
img_name = self.mask[idx].split('/')[-1].split(
'_mask')[0] + '_origin_cut_x1.png'
img = Image.open(self.imgs[img_name])
if self._way == "train":
if np.random.rand() > 0.5:
img = img.transpose(Image.FLIP_LEFT_RIGHT)
mask = mask.transpose(Image.FLIP_LEFT_RIGHT)
if self.use_scale:
img, mask = self._generate_scale_label(np.array(img),
np.array(mask))
img, mask = np.array(img), np.array(mask)
img_h, img_w = mask.shape
pad_h = max(self._crop_size - img_h, 0)
pad_w = max(self._crop_size - img_w, 0)
if pad_h > 0 or pad_w > 0:
img_pad = cv2.copyMakeBorder(img,
0,
pad_h,
0,
pad_w,
cv2.BORDER_CONSTANT,
value=(255.0, 255.0, 255.0))
mask_pad = cv2.copyMakeBorder(mask,
0,
pad_h,
0,
pad_w,
cv2.BORDER_CONSTANT,
value=(0.0, ))
else:
img_pad, mask_pad = img, mask
img_h, img_w = mask_pad.shape
h_off = random.randint(0, img_h - self._crop_size)
w_off = random.randint(0, img_w - self._crop_size)
while (np.all(mask_pad[h_off:h_off + self._crop_size,
w_off:w_off + self._crop_size] == 0)):
h_off = random.randint(0, img_h - self._crop_size)
w_off = random.randint(0, img_w - self._crop_size)
img = np.asarray(img_pad[h_off:h_off + self._crop_size,
w_off:w_off + self._crop_size])
mask = np.asarray(
mask_pad[h_off:h_off + self._crop_size,
w_off:w_off + self._crop_size], np.float32)
img_cp = np.asarray(img, np.uint8)
img, _ = cv2.decolor(img)
img = np.array(img, dtype=np.float32)
img = self.norm_or_Standard(img, self._normalize)
if self._way == "train":
# rotate
num_rotate = np.random.randint(0, 4)
img = np.rot90(img, num_rotate)
mask = np.rot90(mask, num_rotate)
mask = self._mask_transform(mask)
img = torch.from_numpy(img.copy())
img = img.unsqueeze(0)
return img, mask, img_cp, img_name
def paint_text(self, text, i):
""" 使用PIL绘制文本图像，传入画布尺寸，返回文本图像
:param h: 画布高度
:param w: 画布宽度
:return: img
"""
# 创建画布背景
bg_b = np.random.randint(0, 255) # 背景色
bg_g = np.random.randint(0, 255)
bg_r = np.random.randint(0, 255)
# 前景色
fg_b = np.random.randint(0, 255) # 背景色
fg_g = np.random.randint(0, 255)
fg_r = np.random.randint(0, 255)
# 计算前景和背景的彩色相似度
bg_color = sRGBColor(bg_b, bg_g, bg_r)
bg_color = convert_color(bg_color, CMYKColor) # 转cmyk
bg_color = convert_color(bg_color, LabColor)
fg_color = sRGBColor(fg_b, fg_g, fg_r)
fg_color = convert_color(fg_color, CMYKColor) # 转cmyk
fg_color = convert_color(fg_color, LabColor)
delta_e = delta_e_cie2000(bg_color, fg_color)
while delta_e < 150 and delta_e > 250: # 150-250
# 创建画布背景色
bg_b = np.random.randint(0, 255)
bg_g = np.random.randint(0, 255)
bg_r = np.random.randint(0, 255)
# 文字前景色
fg_b = np.random.randint(0, 255)
fg_g = np.random.randint(0, 255)
fg_r = np.random.randint(0, 255)
# 计算前景和背景的彩色相似度
bg_color = sRGBColor(bg_b, bg_g, bg_r)
bg_color = convert_color(bg_color, LabColor)
fg_color = sRGBColor(fg_b, fg_g, fg_r)
fg_color = convert_color(fg_color, LabColor)
delta_e = delta_e_cie2000(bg_color, fg_color)
# 随机选择字体
np.random.shuffle(self.font_name)
cur_fonts = self.fonts.get(self.font_name[0])
keys = list(cur_fonts.keys())
np.random.shuffle(keys)
font = cur_fonts.get(keys[0])
text_size = font.getsize(text)
# 根据字体大小创建画布
img_w = text_size[0]
img_h = text_size[1]
# 文本区域上限
h_space = np.random.randint(6, 20)
w_space = 6
h = img_h + h_space
w = img_w + w_space
canvas = Image.new('RGB', (w, h), (bg_b, bg_g, bg_r))
draw = ImageDraw.Draw(canvas)
# 随机平移
start_x = np.random.randint(2, w_space-2)
start_y = np.random.randint(2, h_space-2)
# 绘制当前文本行
draw.text((start_x, start_y), text, font=font, fill=(fg_b, fg_g, fg_r))
img_array = np.array(canvas)
# 透视失真
src = np.float32([[start_x, start_y],
[start_x + w, start_y],
[start_x + w, start_y + h],
[start_x, start_y + h]])
dst = np.float32([[start_x + np.random.randint(0, 10), start_y + np.random.randint(0, 5)],
[start_x + w - np.random.randint(0, 10), start_y + np.random.randint(0, 5)],
[start_x + w - np.random.randint(0, 10), start_y + h - np.random.randint(0, 5)],
[start_x + np.random.randint(0, 10), start_y + h - np.random.randint(0, 5)]])
M = cv2.getPerspectiveTransform(src, dst)
img_array = cv2.warpPerspective(img_array.copy(), M, (w, h),
borderMode=cv2.BORDER_CONSTANT,
borderValue=(bg_b, bg_g, bg_r))
# Image.fromarray(img_array).show()
# 随机旋转
angle = np.random.randint(-8, 8)
rotated = rotate_bound(img_array, angle=angle, bg_color=(bg_b, bg_g, bg_r))
canvas = Image.fromarray(rotated)
img_array = np.array(canvas.convert('CMYK'))[:,:,0:3] # rgb to cmyk
img_array = cv2.resize(img_array.copy(), (128, 32), interpolation=cv2.INTER_CUBIC) # resize
# noisy = add_noise(img_array)
noisy = img_array
# 统计
# psnr.append(measure.compare_psnr(img_array, noisy))
# ssim.append(measure.compare_ssim(img_array, noisy, multichannel=True))
if not self.color:
# 灰度化
gray = np.zeros_like(img_array) # 初始化灰度图输出矩阵
cv2.decolor(img_array, gray) # opencv以指针方式直接修改
ndimg = Image.fromarray(gray)
# ndimg.show()
# 保存
save_path = os.path.join(self.save_path, '{}.jpeg'.format(i)) # 类别序列即文件名
ndimg.save(save_path)
else:
# 加噪
# 画图看一下
ndimg = Image.fromarray(noisy).convert('CMYK')
# ndimg.show()
# 保存
save_path = os.path.join(self.save_path, '{}.jpeg'.format(i)) # 类别序列即文件名
ndimg.save(save_path)
def PreProcessImage(image):
## TODO searchwindow size based on image size
## Step 0 Non local means denoising
## nonLocalMeans needs nvidia gpu
#x = cv2.cuda.nonLocalMeans(image.copy(),h=7,search_window=35,block_size=7)
#showimage(x,'slownlmeans-0')
## scikit has slower version of nonlocalmeans but running time is very high even for 1mb image
"""
sigma_est = np.mean(estimate_sigma(image.copy(), multichannel=False))
print(f"estimated noise standard deviation = {sigma_est}")
patch_kw = dict(patch_size=7, # 5x5 patches
patch_distance=13, # 35x35 search area
multichannel=False)
# slow algorithm
image = denoise_nl_means(image.copy(), h=0.6 * sigma_est, fast_mode=False, **patch_kw)
showimage(image, 'slow-nlmeans-0')
"""
image = cv2.fastNlMeansDenoisingColored(image,None,h=7,hColor=7,templateWindowSize=7,searchWindowSize=35)
showimage(image, 'nlmeans-0')
##canonical1 Contours 502, good rects 490, compound rects 12, discarded rects 0, chars 427, dropped chars 3
## Contours 504, good rects 493, compound rects 11, discarded rects 0, chars 428, dropped chars 3
##canonical2 Contours 668, good rects 664, compound rects 4, discarded rects 0, chars 562, dropped chars 6
## Contours 625, good rects 623, compound rects 2, discarded rects 0, chars 556, dropped chars 5
## Contours 786, good rects 784, compound rects 1, discarded rects 1, chars 571, dropped chars 1
## Contours 661, good rects 659, compound rects 2, discarded rects 0, chars 567, dropped chars 3
##canonical3 Contours 471, good rects 463, compound rects 8, discarded rects 0, chars 406, dropped chars 2
## Contours 474, good rects 466, compound rects 8, discarded rects 0, chars 406, dropped chars 2
##devnagri/style1 Contours 655, good rects 653, compound rects 0, discarded rects 2, chars 499, dropped chars 1
## Contours 733, good rects 733, compound rects 0, discarded rects 0, chars 515, dropped chars 0
## Contours 671, good rects 671, compound rects 0, discarded rects 0, chars 503, dropped chars 1
## Contours 669, good rects 667, compound rects 0, discarded rects 2, chars 504, dropped chars 1
## Contours 654, good rects 652, compound rects 0, discarded rects 0, chars 500
##tough1 Contours 721, good rects 682, compound rects 34, discarded rects 5, chars 449, dropped chars 22
## Contours 729, good rects 696, compound rects 32, discarded rects 1, chars 514, dropped chars 22
## Contours 687, good rects 648, compound rects 35, discarded rects 4, chars 466, dropped chars 27
## Contours 670, good rects 638, compound rects 30, discarded rects 2, chars 511, dropped chars 23
## Contours 721, good rects 689, compound rects 30, discarded rects 2, chars 513, dropped chars 22
## Contours 721, good rects 689, compound rects 30, discarded rects 2, chars 513, dropped chars 22
## Contours 667, good rects 629, compound rects 36, discarded rects 2, chars 450, dropped chars 45
## Contours 667, good rects 629, compound rects 36, discarded rects 2, chars 454, dropped chars 36
##1238 Contours 793, good rects 788, compound rects 3, discarded rects 2, chars 631, dropped chars 8
##1239 Contours 858, good rects 853, compound rects 5, discarded rects 0, chars 658, dropped chars 21
##1240 Contours 826, good rects 816, compound rects 8, discarded rects 2, chars 663, dropped chars 11
## Step 1 bilateral with conservative params, but multiple times
for i in range(100):
image = cv2.bilateralFilter(image,2,10,10)
showimage(image, 'bilateral-1')
## Step 2 median blur with conservative params
image = cv2.medianBlur(image,3)
showimage(image,'medianblurred-2')
## TODO see if this should be used.
image,_ = cv2.decolor(image.copy())
showimage(image, 'decolor-3')
hist = cv2.calcHist([image],[0],None,[256],[0,256])
plt.plot(hist)
plt.xlabel('decolor')
plt.show()
## Step 3 convert to grey scale
#image = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
#showimage(image, 'grayimage-3')
#hist = cv2.calcHist([image],[0],None,[256],[0,256])
#plt.plot(hist)
#plt.xlabel('regular')
#plt.show()
return image
def convert2(im):
import cv2
import numpy as np
return Image.fromarray(cv2.decolor(np.array(im.convert('RGB')))[0])
