前言
最近写论文需要插入很多图片，为了蒙混过关，找了很多很多数据增强的手段，增强论文的丰富性，大家不要学我哈，反正我把技巧放这儿了！！！哈哈哈哈哈哈哈哈哈
1. Data Augmentation
1.1 数据增强的作用
大家都知道在深度学习网络训练中，模型的样本越充足训练出来的网络模型泛化性越强，鲁棒性越高。最好的例子就是SSD对大目标效果很好，对小目标效果很差，但当使用数据增强后，涨分立马上去了，原因就在于数据增强crop，可以让小目标变成大目标。
- 增加训练的数据量，提高模型的泛化能力
- 增加噪声数据，提升模型的鲁棒性
- 一定程度上能解决过拟合问题，样本过少容易出现过拟合
- 解决样本不平衡问题，例如某个类别过少，数据增强可以增强这个类别的数量。
1.2 图像增强小工具
我写了一款图像增强的小工具，大家可以查看连接下载使用
2. 数据增强的手段
数据增强的手段太多了，我就不一一解释了，直接总结吧
- 几何变换：反转/平移/旋转/缩放/裁剪
- 颜色变换：加噪声/亮度/饱和度/颜色反转/直方图均衡/色彩平衡/
- 基于GAN网络数据增强：用生成模型，生成不同风格的图像
- 高级增强：Cutout/Mixup/One-hot/Cutmix/AutoAugment/Sample pairing/FMix
- 网络中增强：RandAugment/Refinement/Mosaic/DropOut/DropOut/DropBlock
下面简单是介绍一些大家可能不太熟悉的概念：
Cutout：图像中挖掉一个矩形快补0
Mixup：image——两张图片线性叠加，label——one-hot形式进行相同的线性叠加
Cutmix：Cutout + Mixup，image——在图片中挖出来的矩形用另外一张图片填充，label——做同样的叠加
Sample pairing：image——两张图片直接做平均，label——选取其中一张
FMix：采用Fourier变换选取挖掉和补充的区域
推荐一下网站：
https://segmentfault.com/a/1190000022784467
https://jishuin.proginn.com/p/763bfbd29660
https://www.pythonf.cn/read/169722
3. 数据增强的代码
3.1 代码
如果大家想深入了解，建议大家跑跑代码
import os
import cv2
import numpy as np
class Resize(object):
"""
调整大小
"""
def __init__(self, output_size):
self.output_size = output_size
def __call__(self, X, Y):
_X = cv2.resize(X, self.output_size, interpolation=cv2.INTER_NEAREST)
w, h = self.output_size
c = Y.shape[-1]
# _Y = cv2.resize(Y, self.output_size)
_Y = np.zeros((h, w, c))
for i in range(Y.shape[-1]):
_Y[..., i] = cv2.resize(Y[..., i], self.output_size, interpolation=cv2.INTER_NEAREST)
return _X, _Y[...,0]
class Clip(object):
"""
彩色截断
"""
def __init__(self, mini, maxi=None):
if maxi is None:
self.mini, self.maxi = 0, mini
else:
self.mini, self.maxi = mini, maxi
def __call__(self, X, Y):
mini_mask = np.where(X < self.mini)
maxi_mask = np.where(X > self.maxi)
X[mini_mask] = self.mini
X[maxi_mask] = self.maxi
return X, Y
class Normalize(object):
"""
最大最小值归一化
"""
def __init__(self, axis=None):
self.axis = axis
def __call__(self, X, Y):
mini = np.min(X, self.axis)
maxi = np.max(X, self.axis)
X = (X - mini) / (maxi - mini)
X = 255 * X
return X, Y
class Standardize(object):
"""
标准归一化
"""
def __init__(self, axis=None):
self.axis = axis
def __call__(self, X, Y):
mean = np.mean(X, self.axis)
std = np.std(X, self.axis)
X = (X - mean) / std
X = 255 * X
return X, Y
class Flip(object):
"""
反转
"""
def __call__(self, X, Y):
for axis in [0, 1]:
if np.random.rand(1) < 0.5:
X = np.flip(X, axis)
Y = np.flip(Y, axis)
return X, Y
class Crop(object):
def __init__(self, min_size_ratio, max_size_ratio=(1, 1)):
self.min_size_ratio = np.array(list(min_size_ratio))
self.max_size_ratio = np.array(list(max_size_ratio))
def __call__(self, X, Y):
size = np.array(X.shape[:2])
mini = self.min_size_ratio * size
maxi = self.max_size_ratio * size
# random size
h = np.random.randint(mini[0], maxi[0])
w = np.random.randint(mini[1], maxi[1])
# random place
shift_h = np.random.randint(0, size[0] - h)
shift_w = np.random.randint(0, size[1] - w)
X = X[shift_h:shift_h+h, shift_w:shift_w+w]
Y = Y[shift_h:shift_h+h, shift_w:shift_w+w]
return X, Y
class CustomFilter(object):
def __init__(self, kernel):
self.kernel = kernel
def __call__(self, X, Y):
X = cv2.filter2D(X, -1, self.kernel)
return X, Y
class Sharpen(object):
"""
锐化
"""
def __init__(self, max_center=4):
self.max_center = max_center
self.identity = np.array([[0, 0, 0],
[0, 1, 0],
[0, 0, 0]])
self.sharpen = np.array([[ 0, -1, 0],
[-1, 4, -1],
[ 0, -1, 0]]) / 4
def __call__(self, X, Y):
sharp = self.sharpen * np.random.random() * self.max_center
kernel = self.identity + sharp
X = cv2.filter2D(X, -1, kernel)
return X, Y
class GaussianBlur(object):
"""
高斯滤波
"""
def __init__(self, max_kernel=np.array([7, 7])):
self.max_kernel = ((max_kernel + 1) // 2)
def __call__(self, X, Y):
kernel_size = (
np.random.randint(1, self.max_kernel[0]) * 2 + 1,
np.random.randint(1, self.max_kernel[1]) * 2 + 1,
)
X = cv2.GaussianBlur(X, kernel_size, 0)
return X, Y
class Perspective(object):
"""
透视
"""
def __init__(self,
max_ratio_translation=(0.2, 0.2, 0),
max_rotation=(10, 10, 360),
max_scale=(0.1, 0.1, 0.2),
max_shearing=(15, 15, 5)):
self.max_ratio_translation = np.array(max_ratio_translation)
self.max_rotation = np.array(max_rotation)
self.max_scale = np.array(max_scale)
self.max_shearing = np.array(max_shearing)
def __call__(self, X, Y):
# get the height and the width of the image
h, w = X.shape[:2]
max_translation = self.max_ratio_translation * np.array([w, h, 1])
# get the values on each axis
t_x, t_y, t_z = np.random.uniform(-1, 1, 3) * max_translation
r_x, r_y, r_z = np.random.uniform(-1, 1, 3) * self.max_rotation
sc_x, sc_y, sc_z = np.random.uniform(-1, 1, 3) * self.max_scale + 1
sh_x, sh_y, sh_z = np.random.uniform(-1, 1, 3) * self.max_shearing
# convert degree angles to rad
theta_rx = np.deg2rad(r_x)
theta_ry = np.deg2rad(r_y)
theta_rz = np.deg2rad(r_z)
theta_shx = np.deg2rad(sh_x)
theta_shy = np.deg2rad(sh_y)
theta_shz = np.deg2rad(sh_z)
# compute its diagonal
diag = (h ** 2 + w ** 2) ** 0.5
# compute the focal length
f = diag
if np.sin(theta_rz) != 0:
f /= 2 * np.sin(theta_rz)
# set the image from cartesian to projective dimension
H_M = np.array([[1, 0, -w / 2],
[0, 1, -h / 2],
[0, 0, 1],
[0, 0, 1]])
# set the image projective to carrtesian dimension
Hp_M = np.array([[f, 0, w / 2, 0],
[0, f, h / 2, 0],
[0, 0, 1, 0]])
# adjust the translation on z
t_z = (f - t_z) / sc_z ** 2
# translation matrix to translate the image
T_M = np.array([[1, 0, 0, t_x],
[0, 1, 0, t_y],
[0, 0, 1, t_z],
[0, 0, 0, 1]])
# calculate cos and sin of angles
sin_rx, cos_rx = np.sin(theta_rx), np.cos(theta_rx)
sin_ry, cos_ry = np.sin(theta_ry), np.cos(theta_ry)
sin_rz, cos_rz = np.sin(theta_rz), np.cos(theta_rz)
# get the rotation matrix on x axis
R_Mx = np.array([[1, 0, 0, 0],
[0, cos_rx, -sin_rx, 0],
[0, sin_rx, cos_rx, 0],
[0, 0, 0, 1]])
# get the rotation matrix on y axis
R_My = np.array([[cos_ry, 0, -sin_ry, 0],
[ 0, 1, 0, 0],
[sin_ry, 0, cos_ry, 0],
[ 0, 0, 0, 1]])
# get the rotation matrix on z axis
R_Mz = np.array([[cos_rz, -sin_rz, 0, 0],
[sin_rz, cos_rz, 0, 0],
[ 0, 0, 1, 0],
[ 0, 0, 0, 1]])
# compute the full rotation matrix
R_M = np.dot(np.dot(R_Mx, R_My), R_Mz)
# get the scaling matrix
Sc_M = np.array([[sc_x, 0, 0, 0],
[ 0, sc_y, 0, 0],
[ 0, 0, sc_z, 0],
[ 0, 0, 0, 1]])
# get the tan of angles
tan_shx = np.tan(theta_shx)
tan_shy = np.tan(theta_shy)
tan_shz = np.tan(theta_shz)
# get the shearing matrix on x axis
Sh_Mx = np.array([[ 1, 0, 0, 0],
[tan_shy, 1, 0, 0],
[tan_shz, 0, 1, 0],
[ 0, 0, 0, 1]])
# get the shearing matrix on y axis
Sh_My = np.array([[1, tan_shx, 0, 0],
[0, 1, 0, 0],
[0, tan_shz, 1, 0],
[0, 0, 0, 1]])
# get the shearing matrix on z axis
Sh_Mz = np.array([[1, 0, tan_shx, 0],
[0, 1, tan_shy, 0],
[0, 0, 1, 0],
[0, 0, 0, 1]])
# compute the full shearing matrix
Sh_M = np.dot(np.dot(Sh_Mx, Sh_My), Sh_Mz)
Identity = np.array([[1, 0, 0, 0],
[0, 1, 0, 0],
[0, 0, 1, 0],
[0, 0, 0, 1]])
# compute the full transform matrix
M = Identity
M = np.dot(Sh_M, M)
M = np.dot(R_M, M)
M = np.dot(Sc_M, M)
M = np.dot(T_M, M)
M = np.dot(Hp_M, np.dot(M, H_M))
# apply the transformation
X = cv2.warpPerspective(X, M, (w, h))
Y = cv2.warpPerspective(Y, M, (w, h))
return X, Y
class Cutout(object):
"""
随机擦除
"""
def __init__(self,
min_size_ratio,
max_size_ratio,
channel_wise=False,
crop_target=True,
max_crop=10,
replacement=0):
self.min_size_ratio = np.array(list(min_size_ratio))
self.max_size_ratio = np.array(list(max_size_ratio))
self.channel_wise = channel_wise
self.crop_target = crop_target
self.max_crop = max_crop
self.replacement = replacement
def __call__(self, X, Y):
size = np.array(X.shape[:2])
mini = self.min_size_ratio * size
maxi = self.max_size_ratio * size
for _ in range(self.max_crop):
# random size
h = np.random.randint(mini[0], maxi[0])
w = np.random.randint(mini[1], maxi[1])
# random place
shift_h = np.random.randint(0, size[0] - h)
shift_w = np.random.randint(0, size[1] - w)
if self.channel_wise:
c = np.random.randint(0, X.shape[-1])
X[shift_h:shift_h+h, shift_w:shift_w+w, c] = self.replacement
if self.crop_target:
Y[shift_h:shift_h+h, shift_w:shift_w+w] = self.replacement
else:
X[shift_h:shift_h+h, shift_w:shift_w+w] = self.replacement
if self.crop_target:
Y[shift_h:shift_h+h, shift_w:shift_w+w] = self.replacement
return X, Y
class Leaf(object):
def __init__(self):
pass
def __call__(self, X, Y):
blur = cv2.GaussianBlur(X, (7, 7), 0)
hsv_blur = cv2.cvtColor(blur, cv2.COLOR_BGR2HSV)
# lower mask (0-10)
lower_red = np.array([0,130,130])
upper_red = np.array([20,255,255])
mask_0 = cv2.inRange(hsv_blur, lower_red, upper_red)
# upper mask (170-180)
lower_red = np.array([165,130,130])
upper_red = np.array([185,255,255])
mask_1 = cv2.inRange(hsv_blur, lower_red, upper_red)
hsv_blur[np.where(mask_1)] = hsv_blur[np.where(mask_1)] - np.array([165, 0, 0])
mask = mask_0 + mask_1
# change color
turn_color = np.random.randint(0, 255)
hsv_blur[np.where(mask)] = hsv_blur[np.where(mask)] + np.array([turn_color, 0, 0])
X_blur = cv2.cvtColor(hsv_blur, cv2.COLOR_HSV2BGR)
X[np.where(mask)] = X_blur[np.where(mask)]
return X, Y
class Brightness(object):
"""
随机改变图像的亮度
"""
def __init__(self, range_brightness=(-50, 50)):
self.range_brightness = range_brightness
def __call__(self, X, Y):
brightness = np.random.randint(*self.range_brightness)
X = X + brightness
return X, Y
class Contrast(object):
"""
改变图像的对比度
"""
def __init__(self, range_contrast=(-50, 50)):
self.range_contrast = range_contrast
def __call__(self, X, Y):
contrast = np.random.randint(*self.range_contrast)
X = X * (contrast / 127 + 1) - contrast
return X, Y
class UniformNoise(object):
"""
均匀噪声
"""
def __init__(self, low=-50, high=50):
self.low = low
self.high = high
def __call__(self, X, Y):
noise = np.random.uniform(self.low, self.high, X.shape)
X = X + noise
return X, Y
class GaussianNoise(object):
def __init__(self, center=0, std=50):
self.center = center
self.std = std
def __call__(self, X, Y):
noise = np.random.normal(self.center, self.std, X.shape)
X = X + noise
return X, Y
class Vignetting(object):
def __init__(self,
ratio_min_dist=0.2,
range_vignette=(0.2, 0.8),
random_sign=False):
self.ratio_min_dist = ratio_min_dist
self.range_vignette = np.array(range_vignette)
self.random_sign = random_sign
def __call__(self, X, Y):
h, w = X.shape[:2]
min_dist = np.array([h, w]) / 2 * np.random.random() * self.ratio_min_dist
# create matrix of distance from the center on the two axis
x, y = np.meshgrid(np.linspace(-w/2, w/2, w), np.linspace(-h/2, h/2, h))
x, y = np.abs(x), np.abs(y)
# create the vignette mask on the two axis
x = (x - min_dist[0]) / (np.max(x) - min_dist[0])
x = np.clip(x, 0, 1)
y = (y - min_dist[1]) / (np.max(y) - min_dist[1])
y = np.clip(y, 0, 1)
# then get a random intensity of the vignette
vignette = (x + y) / 2 * np.random.uniform(*self.range_vignette)
vignette = np.tile(vignette[..., None], [1, 1, 3])
sign = 2 * (np.random.random() < 0.5) * (self.random_sign) - 1
X = X * (1 + sign * vignette)
return X, Y
class LensDistortion(object):
def __init__(self, d_coef=(0.15, 0.15, 0.1, 0.1, 0.05)):
self.d_coef = np.array(d_coef)
def __call__(self, X, Y):
# get the height and the width of the image
h, w = X.shape[:2]
# compute its diagonal
f = (h ** 2 + w ** 2) ** 0.5
# set the image projective to carrtesian dimension
K = np.array([[f, 0, w / 2],
[0, f, h / 2],
[0, 0, 1]])
d_coef = self.d_coef * np.random.random(5) # value
d_coef = d_coef * (2 * (np.random.random(5) < 0.5) - 1) # sign
# Generate new camera matrix from parameters
M, _ = cv2.getOptimalNewCameraMatrix(K, d_coef, (w, h), 0)
# Generate look-up tables for remapping the camera image
remap = cv2.initUndistortRectifyMap(K, d_coef, None, M, (w, h), 5)
# Remap the original image to a new image
X = cv2.remap(X, *remap, cv2.INTER_LINEAR)
Y = cv2.remap(Y, *remap, cv2.INTER_LINEAR)
return X, Y
def data_aug(img_path, label_path, save_img, save_label):
imgs = os.listdir(img_path)
for im_name in imgs:
name = im_name.split('.')[0]
im_full_path = os.path.join(img_path, im_name)
lab_full_path = os.path.join(label_path, im_name)
# im_arr = cv2.imread(im_full_path)
# lab_arr = cv2.imread(lab_full_path)
# save_img_path = os.path.join(save_img, name + '_Resize960.png')
# save_lab_path = os.path.join(save_label, name + '_Resize960.png')
# im_Resize, lab_Resize = Resize((960,960))(im_arr, lab_arr)
# cv2.imwrite(save_img_path, im_Resize, [int(cv2.IMWRITE_PNG_COMPRESSION),0])
# cv2.imwrite(save_lab_path, lab_Resize, [int(cv2.IMWRITE_PNG_COMPRESSION),0])
# im_arr = cv2.imread(im_full_path)
# lab_arr = cv2.imread(lab_full_path)
# save_img_path = os.path.join(save_img, name + '_Resize384.png')
# save_lab_path = os.path.join(save_label, name + '_Resize384.png')
# im_Resize, lab_Resize = Resize((384,384))(im_arr, lab_arr)
# cv2.imwrite(save_img_path, im_Resize, [int(cv2.IMWRITE_PNG_COMPRESSION),0])
# cv2.imwrite(save_lab_path, lab_Resize, [int(cv2.IMWRITE_PNG_COMPRESSION),0])
# im_arr = cv2.imread(im_full_path)
# lab_arr = cv2.imread(lab_full_path,0)
# save_img_path = os.path.join(save_img, name + '_Clip.png')
# save_lab_path = os.path.join(save_label, name + '_Clip.png')
# im_Clip, lab_Clip = Clip(20,150)(im_arr, lab_arr)
# cv2.imwrite(save_img_path, im_Clip, [int(cv2.IMWRITE_PNG_COMPRESSION),0])
# cv2.imwrite(save_lab_path, lab_Clip, [int(cv2.IMWRITE_PNG_COMPRESSION),0])
# im_arr = cv2.imread(im_full_path)
# lab_arr = cv2.imread(lab_full_path,0)
# save_img_path = os.path.join(save_img, name + '_Normalize.png')
# save_lab_path = os.path.join(save_label, name + '_Normalize.png')
# im_Normalize, lab_Normalize = Normalize(1)(im_arr, lab_arr)
# cv2.imwrite(save_img_path, im_Normalize, [int(cv2.IMWRITE_PNG_COMPRESSION),0])
# cv2.imwrite(save_lab_path, lab_Normalize, [int(cv2.IMWRITE_PNG_COMPRESSION),0])
# im_arr = cv2.imread(im_full_path)
# lab_arr = cv2.imread(lab_full_path,0)
# save_img_path = os.path.join(save_img, name + '_Standardize.png')
# save_lab_path = os.path.join(save_label, name + '_Standardize.png')
# im_Standardize, lab_Standardize = Standardize(1)(im_arr, lab_arr)
# cv2.imwrite(save_img_path, im_Standardize, [int(cv2.IMWRITE_PNG_COMPRESSION),0])
# cv2.imwrite(save_lab_path, lab_Standardize, [int(cv2.IMWRITE_PNG_COMPRESSION),0])
# im_arr = cv2.imread(im_full_path)
# lab_arr = cv2.imread(lab_full_path)
# save_img_path = os.path.join(save_img, name + '_Flip.png')
# save_lab_path = os.path.join(save_label, name + '_Flip.png')
# im__Flip, lab__Flip= Flip()(im_arr, lab_arr)
# cv2.imwrite(save_img_path, im__Flip, [int(cv2.IMWRITE_PNG_COMPRESSION),0])
# cv2.imwrite(save_lab_path, lab__Flip, [int(cv2.IMWRITE_PNG_COMPRESSION),0])
# im_arr = cv2.imread(im_full_path)
# lab_arr = cv2.imread(lab_full_path,0)
# save_img_path = os.path.join(save_img, name + '_Crop.png')
# save_lab_path = os.path.join(save_label, name + '_Crop.png')
# im__Crop, lab__Crop = Crop((0.5,0.5))(im_arr, lab_arr)
# cv2.imwrite(save_img_path, im__Crop, [int(cv2.IMWRITE_PNG_COMPRESSION),0])
# cv2.imwrite(save_lab_path, lab__Crop, [int(cv2.IMWRITE_PNG_COMPRESSION),0])
# im_arr = cv2.imread(im_full_path)
# lab_arr = cv2.imread(lab_full_path,0)
# save_img_path = os.path.join(save_img, name + '_CustomFilter.png')
# save_lab_path = os.path.join(save_label, name + '_CustomFilter.png')
# lpls = np.array([[-1,1,-1],[1,8,-1],[-1,1,-1]])
# im_CustomFilter, lab_CustomFilter = CustomFilter(lpls)(im_arr, lab_arr)
# cv2.imwrite(save_img_path, im_CustomFilter, [int(cv2.IMWRITE_PNG_COMPRESSION),0])
# cv2.imwrite(save_lab_path, lab_CustomFilter, [int(cv2.IMWRITE_PNG_COMPRESSION),0])
# im_arr = cv2.imread(im_full_path)
# lab_arr = cv2.imread(lab_full_path,0)
# save_img_path = os.path.join(save_img, name + '_Sharpen.png')
# save_lab_path = os.path.join(save_label, name + '_Sharpen.png')
# im__Sharpen, lab__Sharpen = Sharpen()(im_arr, lab_arr)
# cv2.imwrite(save_img_path, im__Sharpen, [int(cv2.IMWRITE_PNG_COMPRESSION),0])
# cv2.imwrite(save_lab_path, lab__Sharpen, [int(cv2.IMWRITE_PNG_COMPRESSION),0])
# im_arr = cv2.imread(im_full_path)
# lab_arr = cv2.imread(lab_full_path,0)
# save_img_path = os.path.join(save_img, name + '_GaussianBlur.png')
# save_lab_path = os.path.join(save_label, name + '_GaussianBlur.png')
# im__GaussianBlur, lab__GaussianBlur = GaussianBlur()(im_arr, lab_arr)
# cv2.imwrite(save_img_path, im__GaussianBlur, [int(cv2.IMWRITE_PNG_COMPRESSION),0])
# cv2.imwrite(save_lab_path, lab__GaussianBlur, [int(cv2.IMWRITE_PNG_COMPRESSION),0])
# im_arr = cv2.imread(im_full_path)
# lab_arr = cv2.imread(lab_full_path,0)
# save_img_path = os.path.join(save_img, name + '_Perspective.png')
# save_lab_path = os.path.join(save_label, name + '_Perspective.png')
# im__Perspective, lab__Perspective = Perspective()(im_arr, lab_arr)
# cv2.imwrite(save_img_path, im__Perspective, [int(cv2.IMWRITE_PNG_COMPRESSION),0])
# cv2.imwrite(save_lab_path, lab__Perspective, [int(cv2.IMWRITE_PNG_COMPRESSION),0])
# im_arr = cv2.imread(im_full_path)
# lab_arr = cv2.imread(lab_full_path,0)
# save_img_path = os.path.join(save_img, name + '_Cutout.png')
# save_lab_path = os.path.join(save_label, name + '_Cutout.png')
# im__Cutout, lab__Cutout = Cutout([0.01,0.05],[0.2,0.3])(im_arr, lab_arr)
# cv2.imwrite(save_img_path, im__Cutout, [int(cv2.IMWRITE_PNG_COMPRESSION),0])
# cv2.imwrite(save_lab_path, lab__Cutout, [int(cv2.IMWRITE_PNG_COMPRESSION),0])
# im_arr = cv2.imread(im_full_path)
# lab_arr = cv2.imread(lab_full_path,0)
# save_img_path = os.path.join(save_img, name + '_Leaf.png')
# save_lab_path = os.path.join(save_label, name + '_Leaf.png')
# im__Leaf, lab__Leaf = Leaf()(im_arr, lab_arr)
# cv2.imwrite(save_img_path, im__Leaf, [int(cv2.IMWRITE_PNG_COMPRESSION),0])
# cv2.imwrite(save_lab_path, lab__Leaf, [int(cv2.IMWRITE_PNG_COMPRESSION),0])
# im_arr = cv2.imread(im_full_path)
# lab_arr = cv2.imread(lab_full_path,0)
# save_img_path = os.path.join(save_img, name + '_Brightness.png')
# save_lab_path = os.path.join(save_label, name + '_Brightness.png')
# im__Brightness, lab__Brightness = Brightness()(im_arr, lab_arr)
# cv2.imwrite(save_img_path, im__Brightness, [int(cv2.IMWRITE_PNG_COMPRESSION),0])
# cv2.imwrite(save_lab_path, lab__Brightness, [int(cv2.IMWRITE_PNG_COMPRESSION),0])
# im_arr = cv2.imread(im_full_path)
# lab_arr = cv2.imread(lab_full_path,0)
# save_img_path = os.path.join(save_img, name + '_Contrast.png')
# save_lab_path = os.path.join(save_label, name + '_Contrast.png')
# im__Contrast, lab__Contrast = Contrast()(im_arr, lab_arr)
# cv2.imwrite(save_img_path, im__Contrast, [int(cv2.IMWRITE_PNG_COMPRESSION),0])
# cv2.imwrite(save_lab_path, lab__Contrast, [int(cv2.IMWRITE_PNG_COMPRESSION),0])
# im_arr = cv2.imread(im_full_path)
# lab_arr = cv2.imread(lab_full_path,0)
# save_img_path = os.path.join(save_img, name + '_UniformNoise.png')
# save_lab_path = os.path.join(save_label, name + '_UniformNoise.png')
# im__UniformNoise, lab__UniformNoise = UniformNoise()(im_arr, lab_arr)
# cv2.imwrite(save_img_path, im__UniformNoise, [int(cv2.IMWRITE_PNG_COMPRESSION),0])
# cv2.imwrite(save_lab_path, lab__UniformNoise, [int(cv2.IMWRITE_PNG_COMPRESSION),0])
# im_arr = cv2.imread(im_full_path)
# lab_arr = cv2.imread(lab_full_path,0)
# save_img_path = os.path.join(save_img, name + '_GaussianNoise.png')
# save_lab_path = os.path.join(save_label, name + '_GaussianNoise.png')
# im__GaussianNoise, lab__GaussianNoise = GaussianNoise()(im_arr, lab_arr)
# cv2.imwrite(save_img_path, im__GaussianNoise, [int(cv2.IMWRITE_PNG_COMPRESSION),0])
# cv2.imwrite(save_lab_path, lab__GaussianNoise, [int(cv2.IMWRITE_PNG_COMPRESSION),0])
# im_arr = cv2.imread(im_full_path)
# lab_arr = cv2.imread(lab_full_path,0)
# save_img_path = os.path.join(save_img, name + '_Vignetting.png')
# save_lab_path = os.path.join(save_label, name + '_Vignetting.png')
# im__Vignetting, lab__Vignetting = Vignetting()(im_arr, lab_arr)
# cv2.imwrite(save_img_path, im__Vignetting, [int(cv2.IMWRITE_PNG_COMPRESSION),0])
# cv2.imwrite(save_lab_path, lab__Vignetting, [int(cv2.IMWRITE_PNG_COMPRESSION),0])
im_arr = cv2.imread(im_full_path)
lab_arr = cv2.imread(lab_full_path,0)
save_img_path = os.path.join(save_img, name + '_LensDistortion.png')
save_lab_path = os.path.join(save_label, name + '_LensDistortion.png')
im__LensDistortion, lab__LensDistortion = LensDistortion()(im_arr, lab_arr)
cv2.imwrite(save_img_path, im__LensDistortion, [int(cv2.IMWRITE_PNG_COMPRESSION),0])
cv2.imwrite(save_lab_path, lab__LensDistortion, [int(cv2.IMWRITE_PNG_COMPRESSION),0])
if __name__ == "__main__":
img_path = './crop/1/'
label_path = './crop/1/'
save_img = './crop/1/'
if not os.path.exists(save_img):
os.makedirs(save_img)
save_label = './crop/1/'
if not os.path.exists(save_label):
os.makedirs(save_label)
data_aug(img_path, label_path, save_img, save_label)
3.2 结果
对了，建议大家用彩色图片，效果明显，我直接用我论文的图了。
a）Image；b）Resize；c）Clip；d）Normalize；f）Standardize；g）Flip；h）Crop i）Custom Filter；j）Sharpen；k）Gaussian Blur；l）Perspective；m）Cutout；n）Brightness；o）Contrast；p）Uniform Noise；q）Gaussian Noise文章来源：https://www.toymoban.com/news/detail-719510.html
代码2文章来源地址https://www.toymoban.com/news/detail-719510.html
import torchvision.transforms as transforms
from PIL import Image
img = Image.open('tina.jpg')
'''
# CenterCrop
size = (224, 224)
transform = transforms.CenterCrop(size)
center_crop = transform(img)
center_crop.save('center_crop.jpg')
# ColorJitter
brightness = (1, 10)
contrast = (1, 10)
saturation = (1, 10)
hue = (0.2, 0.4)
transform = transforms.ColorJitter(brightness, contrast, saturation, hue)
color_jitter = transform(img)
color_jitter.save('color_jitter.jpg')
# FiveCrop
size = (224, 224)
transform = transforms.FiveCrop(size)
five_crop = transform(img)
for index, img in enumerate(five_crop):
img.save(str(index) + '.jpg')
# Grayscale
transform = transforms.Grayscale()
grayscale = transform(img)
grayscale.save('grayscale.jpg')
# Compose, Pad
size = (224, 224)
padding = 16
fill = (0, 0, 255)
transform = transforms.Compose([
transforms.CenterCrop(size),
transforms.Pad(padding, fill)
])
pad = transform(img)
pad.save('pad.jpg')
# RandomAffine
degrees = (15, 30)
translate=(0, 0.2)
scale=(0.8, 1)
fillcolor = (0, 0, 255)
transform = transforms.RandomAffine(degrees=degrees, translate=translate, scale=scale, fillcolor=fillcolor)
random_affine = transform(img)
random_affine.save('random_affine.jpg')
# RandomApply
size = (224, 224)
padding = 16
fill = (0, 0, 255)
transform = transforms.RandomApply([transforms.CenterCrop(size), transforms.Pad(padding, fill)])
for i in range(3):
random_apply = transform(img)
random_apply.save(str(i) + '.jpg')
# RandomChoice
transform = transforms.RandomChoice([transforms.RandomAffine(degrees),
transforms.CenterCrop(size),
transforms.Pad(padding, fill)])
for i in range(3):
random_order = transform(img)
random_order.save(str(i) + '.jpg')
# RandomCrop
size = (224, 224)
transform = transforms.RandomCrop(size)
random_crop = transform(img)
random_crop.save('p.jpg')
# RandomGrayscale
p = 0.5
transform = transforms.RandomGrayscale(p)
for i in range(3):
random_grayscale = transform(img)
random_grayscale.save(str(i) + '.jpg')
# RandomHorizontalFlip
p = 0.5
transform = transforms.RandomHorizontalFlip(p)
for i in range(3):
random_horizontal_filp = transform(img)
random_horizontal_filp.save(str(i) + '.jpg')
# RandomOrder
size = (224, 224)
padding = 16
fill = (0, 0, 255)
degrees = (15, 30)
transform = transforms.RandomOrder([transforms.RandomAffine(degrees),
transforms.CenterCrop(size),
transforms.Pad(padding, fill)])
for i in range(3):
random_order = transform(img)
random_order.save(str(i) + '.jpg')
# RandomPerspective
distortion_scale = 1
p = 1
fill = (0, 0, 255)
transform = transforms.RandomPerspective(distortion_scale=distortion_scale, p=p, fill=fill)
random_perspective = transform(img)
random_perspective.save('random_perspective.jpg')
# RandomResizedCrop
size = (256, 256)
scale=(0.8, 1.0)
ratio=(0.75, 1.0)
transform = transforms.RandomResizedCrop(size=size, scale=scale, ratio=ratio)
random_resized_crop = transform(img)
random_resized_crop.save('random_resized_crop.jpg')
# RandomRotation
degrees = (15, 30)
fill = (0, 0, 255)
transform = transforms.RandomRotation(degrees=degrees, fill=fill)
random_rotation = transform(img)
random_rotation.save('random_rotation.jpg')
# RandomVerticalFlip
p = 1
transform = transforms.RandomVerticalFlip(p)
random_vertical_filp = transform(img)
random_vertical_filp.save('random_vertical_filp.jpg')
# Resize
size = (224, 224)
transform = transforms.Resize(size)
resize_img = transform(img)
resize_img.save('resize_img.jpg')
# ToPILImage
img = Image.open('tina.jpg')
transform = transforms.ToTensor()
img = transform(img)
print(img.size())
img_r = img[0, :, :]
img_g = img[1, :, :]
img_b = img[2, :, :]
print(type(img_r))
print(img_r.size())
transform = transforms.ToPILImage()
img_r = transform(img_r)
img_g = transform(img_g)
img_b = transform(img_b)
print(type(img_r))
img_r.save('img_r.jpg')
img_g.save('img_g.jpg')
img_b.save('img_b.jpg')
# ToTensor
img = Image.open('tina.jpg')
print(type(img))
print(img.size)
transform = transforms.ToTensor()
img = transform(img)
print(type(img))
print(img.size())
'''
# RandomErasing
p = 1.0
scale = (0.2, 0.3)
ratio = (0.5, 1.0)
value = (0, 0, 255)
transform = transforms.Compose([
transforms.ToTensor(),
transforms.RandomErasing(p=p, scale=scale, ratio=ratio, value=value),
transforms.ToPILImage()
])
random_erasing = transform(img)
random_erasing.save('random_erasing.jpg')
到了这里，关于数据增强(Data Augmentation)的文章就介绍完了。如果您还想了解更多内容，请在右上角搜索TOY模板网以前的文章或继续浏览下面的相关文章，希望大家以后多多支持TOY模板网！
