数据增强策略：
1 在线模式--训练中
随机裁剪(完全随机，四个角+中心) crop
def random_crop(img, scale=[0.8, 1.0], ratio=[3. / 4., 4. / 3.], resize_w=100, resize_h=100):
"""
随机裁剪
:param img:
:param scale: 缩放
:param ratio:
:param resize_w:
:param resize_h:
:return:
"""
aspect_ratio = math.sqrt(np.random.uniform(*ratio))
w = 1. * aspect_ratio
h = 1. / aspect_ratio
src_h, src_w = img.shape[:2]
bound = min((float(src_w) / src_h) / (w ** 2),
(float(src_h) / src_w) / (h ** 2))
scale_max = min(scale[1], bound)
scale_min = min(scale[0], bound)
target_area = src_h * src_w * np.random.uniform(scale_min,
scale_max)
target_size = math.sqrt(target_area)
w = int(target_size * w)
h = int(target_size * h)
i = np.random.randint(0, src_w - w + 1)
j = np.random.randint(0, src_h - h + 1)
img = img[j:j + h, i:i + w]
img = cv2.resize(img, (resize_w, resize_h))
return img
def rule_crop(img, box_ratio=(3. / 4, 3. / 4), location_type='LT', resize_w=100, resize_h=100):
"""
按照一定规则进行裁剪, 直接在原图尺寸上操作，不对原图进行
:param img:
:param box_ratio: 剪切的 比例： （宽度上的比例， 高度上的比例）
:param location_type: 具体在=哪个位置： 以下其中一个：
LR : 左上角
RT : 右上角
LB : 左下角
RB : 右下角
CC : 中心
:param resize_w: 输出图的width
:param resize_h: 输出图的height
:return:
"""
assert location_type in ('LT', 'RT', 'LB', 'RB', 'CC'), 'must have a location .'
is_gray = False
if len(img.shape) == 3:
h, w, c = img.shape
elif len(img.shape) == 2:
h, w = img.shape
is_gray = True
crop_w, crop_h = int(w * box_ratio[0]), int(h * box_ratio[1])
crop_img = np.zeros([10, 10])
if location_type == 'LT':
crop_img = img[:crop_h, :crop_w, :] if not is_gray else img[:crop_h, :crop_w]
elif location_type == 'RT':
crop_img = img[:crop_h:, w - crop_w:, :] if not is_gray else img[:crop_h:, w - crop_w:]
elif location_type == 'LB':
crop_img = img[h - crop_h:, :crop_w, :] if not is_gray else img[h - crop_h:, :crop_w]
elif location_type == 'RB':
crop_img = img[h - crop_h:, w - crop_w:, :] if not is_gray else img[h - crop_h:, w - crop_w:]
elif location_type == 'CC':
start_h = (h - crop_h) // 2
start_w = (w - crop_w) // 2
crop_img = img[start_h:start_h + crop_h, start_w:start_w + crop_w, :] if not is_gray else img[
start_h:start_h + crop_h,
start_w:start_w + crop_w]
resize = cv2.resize(crop_img, (resize_w, resize_h))
return resize
水平翻转 flip
def random_flip(img, mode=1):
"""
随机翻转
:param img:
:param model: 1=水平翻转 / 0=垂直 / -1=水平垂直
:return:
"""
assert mode in (0, 1, -1), "mode is not right"
flip = np.random.choice(2) * 2 - 1 # -1 / 1
if mode == 1:
img = img[:, ::flip, :]
elif mode == 0:
img = img[::flip, :, :]
elif mode == -1:
img = img[::flip, ::flip, :]
return img
def flip(img, mode=1):
"""
翻转
:param img:
:param mode: 1=水平翻转 / 0=垂直 / -1=水平垂直
:return:
"""
assert mode in (0, 1, -1), "mode is not right"
return cv2.flip(img, flipCode=mode)
2 离线模式
2.1 随机扰动
噪声(高斯、自定义) noise
def random_noise(img, rand_range=(3, 20)):
"""
随机噪声
:param img:
:param rand_range: (min, max)
:return:
"""
img = np.asarray(img, np.float)
sigma = random.randint(*rand_range)
nosie = np.random.normal(0, sigma, size=img.shape)
img += nosie
img = np.uint8(np.clip(img, 0, 255))
return img
滤波(高斯、平滑、均值、中值、最大最小值、双边、引导、运动)
# 各种滤波原理介绍：https://blog.csdn.net/hellocsz/article/details/80727972
def gaussianBlue(img, ks=(7, 7), stdev=1.5):
"""
高斯模糊, 可以对图像进行平滑处理，去除尖锐噪声
:param img:
:param ks: 卷积核
:param stdev: 标准差
:return:
"""
return cv2.GaussianBlur(img, (7, 7), 1.5)
2.2 转换
旋转 rorate
def rotate(img, angle, scale=1.0):
"""
旋转
:param img:
:param angle: 旋转角度， >0 表示逆时针，
:param scale:
:return:
"""
height, width = img.shape[:2] # 获取图像的高和宽
center = (width / 2, height / 2) # 取图像的中点
M = cv2.getRotationMatrix2D(center, angle, scale) # 获得图像绕着某一点的旋转矩阵
# cv2.warpAffine()的第二个参数是变换矩阵,第三个参数是输出图像的大小
rotated = cv2.warpAffine(img, M, (height, width))
return rotated
def random_rotate(img, angle_range=(-10, 10)):
"""
随机旋转
:param img:
:param angle_range: 旋转角度范围 (min,max) >0 表示逆时针，
:return:
"""
height, width = img.shape[:2] # 获取图像的高和宽
center = (width / 2, height / 2) # 取图像的中点
angle = random.randrange(*angle_range, 1)
M = cv2.getRotationMatrix2D(center, angle, 1.0) # 获得图像绕着某一点的旋转矩阵
# cv2.warpAffine()的第二个参数是变换矩阵,第三个参数是输出图像的大小
rotated = cv2.warpAffine(img, M, (height, width))
return rotated
偏移 shift
def shift(img, x_offset, y_offset):
"""
偏移，向右 向下
:param img:
:param x_offset: >0表示向右偏移px, <0表示向左
:param y_offset: >0表示向下偏移px, <0表示向上
:return:
"""
h, w, _ = img.shape
M = np.array([[1, 0, x_offset], [0, 1, y_offset]], dtype=np.float)
return cv2.warpAffine(img, M, (w, h))
扭曲 skew
...
缩放 scale
def resize_img(img, resize_w, resize_h):
height, width = img.shape[:2] # 获取图片的高和宽
return cv2.resize(img, (resize_w, resize_h), interpolation=cv2.INTER_CUBIC)
RGB/BGR->HSV
def rgb2hsv_py(r, g, b):
# from https://blog.csdn.net/weixin_43360384/article/details/84871521
r, g, b = r/255.0, g/255.0, b/255.0
mx = max(r, g, b)
mn = min(r, g, b)
m = mx-mn
if mx == mn:
h = 0
elif mx == r:
if g >= b:
h = ((g-b)/m)*60
else:
h = ((g-b)/m)*60 + 360
elif mx == g:
h = ((b-r)/m)*60 + 120
elif mx == b:
h = ((r-g)/m)*60 + 240
if mx == 0:
s = 0
else:
s = m/mx
v = mx
return h, s, v
def rgb2hsv_cv(img):
# from https://blog.csdn.net/qq_38332453/article/details/89258058
h = img.shape[0]
w = img.shape[1]
H = np.zeros((h,w),np.float32)
S = np.zeros((h, w), np.float32)
V = np.zeros((h, w), np.float32)
r,g,b = cv2.split(img)
r, g, b = r/255.0, g/255.0, b/255.0
for i in range(0, h):
for j in range(0, w):
mx = max((b[i, j], g[i, j], r[i, j]))
mn = min((b[i, j], g[i, j], r[i, j]))
V[i, j] = mx
if V[i, j] == 0:
S[i, j] = 0
else:
S[i, j] = (V[i, j] - mn) / V[i, j]
if mx == mn:
H[i, j] = 0
elif V[i, j] == r[i, j]:
if g[i, j] >= b[i, j]:
H[i, j] = (60 * ((g[i, j]) - b[i, j]) / (V[i, j] - mn))
else:
H[i, j] = (60 * ((g[i, j]) - b[i, j]) / (V[i, j] - mn))+360
elif V[i, j] == g[i, j]:
H[i, j] = 60 * ((b[i, j]) - r[i, j]) / (V[i, j] - mn) + 120
elif V[i, j] == b[i, j]:
H[i, j] = 60 * ((r[i, j]) - g[i, j]) / (V[i, j] - mn) + 240
H[i,j] = H[i,j] / 2
return H, S, V
图片叠加与融合
def addWeight(src1, alpha, src2, beta, gamma):
"""
g (x) = (1 − α)f0 (x) + αf1 (x) #a→（0，1）不同的a值可以实现不同的效果
dst = src1 * alpha + src2 * beta + gamma
:param src1: img1
:param alpha:
:param src2: img2
:param beta:
:param gamma:
:return:
"""
assert src1.shap == src2.shape
return cv2.addWeighted(src1, alpha, src2, beta, gamma)
颜色抖动(亮度\色度\饱和度\对比度) color jitter
def adjust_contrast_bright(img, contrast=1.2, brightness=100):
"""
调整亮度与对比度
dst = img * contrast + brightness
:param img:
:param contrast: 对比度 越大越亮
:param brightness: 亮度 0~100
:return:
"""
# 像素值会超过0-255， 因此需要截断
return np.uint8(np.clip((contrast * img + brightness), 0, 255))
def pytorch_color_jitter(img):
return torchvision.transforms.ColorJitter(brightness=0, contrast=0, saturation=0, hue=0)
3D几何变换