图片 Augmentation整理
2024-09-05 07:18:53
目录
Augmentation
原图:
Flipping 翻转
Grayscale
Equalize 均衡直方图
Posterize 减少颜色通道位数
Cropping
Rotation
Translation
Noise injection
Hue 色调
Brightness
Saturation
Contrast
Erasing 抹去
Augmix
代码
config.json:
{
"Crop": {
"size": [100, 200]
},
"Rotate": {
"degrees": 20
},
"Translate": {
"translate": [0.3, 0.3]
},
"Scale": {
"scale": [0.7, 1.3]
},
"Shear": {
"shear": [10, 100, 10, 100]
},
"Hflip": {},
"Vflip": {},
"Equalize": {
"mask": null
},
"Posterize": {
"bits" : 2
},
"Grayscale": {
"num_output_channels": 1
},
"Normalnoise": {
"loc": 0.0,
"scale": 0.1
},
"Uniformnoise": {
"alpha": 0.2
},
"Contrast": {
"contrast": 0.8
},
"Brightness": {
"brightness": 0.7
},
"Saturation": {
"saturation": 0.7
},
"Hue": {
"hue": 0.5
},
"Erase": {
"value": 100
}
}
"""
ops.py
"""
import torch
import torchvision.transforms.functional as F
import PIL.ImageOps
import random
from torchvision import transforms
import scipy.stats
from collections.abc import Iterable
def special(cls, name, *args, **kwargs):
class Newclass(cls):
def __init__(self, *newargs, **newkwargs):
newargs += args
newkwargs.update(kwargs)
super(Newclass, self).__init__(*newargs, **newkwargs)
Newclass.__name__ = name
return Newclass
class Hflip:
"""
水平翻转
"""
def __call__(self, img):
return F.hflip(img)
class Vflip:
"""
上下翻转
"""
def __call__(self, img):
return F.vflip(img)
class Equalize:
"""
均衡图像直方图
"""
def __init__(self, mask):
self.mask = mask
def __call__(self, img):
return PIL.ImageOps.equalize(img, self.mask)
class Posterize:
"""
减少每个颜色通道的位数
"""
def __init__(self, bits):
self.bits = bits
def __call__(self, img):
return PIL.ImageOps.posterize(img, self.bits)
class Grayscale:
def __init__(self, num_output_channels=1):
self.num_output_channels = num_output_channels
def __call__(self, img):
return F.to_grayscale(img, self.num_output_channels)
class Normalnoise:
def __init__(self, loc, scale):
self.loc = loc
self.scale = scale
def __call__(self, img):
img = F.to_tensor(img)
img += (torch.randn_like(img) + self.loc) * self.scale
img = torch.clamp(img, 0., 1.)
return F.to_pil_image(img)
class Uniformnoise:
def __init__(self, alpha):
if alpha < 0 or alpha > 1:
raise ValueError("alpha should be in [0, 1]...")
self.alpha = alpha
def __call__(self, img):
img = F.to_tensor(img)
img += torch.rand_like(img) * self.alpha
img = torch.clamp(img, 0., 1.)
return F.to_pil_image(img)
class Erase:
def __init__(self, value):
self.value = value
def __call__(self, img):
img = F.to_tensor(img)
H, W = img.size()[-2:]
lui = torch.randint(0, H, (1,)).item()
luj = torch.randint(0, W, (1,)).item()
rbi = torch.randint(lui, H, (1,)).item()
rbj = torch.randint(luj, W, (1,)).item()
h = rbj - luj
w = rbi - lui
return F.to_pil_image(F.erase(img, lui, luj, h, w, 0))
class Perspective:
def __init__(self, startpoints, endpoints, interpolation=3):
self.startpoints = startpoints
self.endpoints = endpoints
self.interpolation = interpolation
def __call__(self, img):
return F.perspective(img, self.startpoints,
self.endpoints, self.interpolation)
Translate = special(transforms.RandomAffine, "Translate", degrees = 0)
Scale = special(transforms.RandomAffine, "Scale", degrees = 0)
Shear = special(transforms.RandomAffine, "Shear", degrees=0)
Rotate = special(transforms.RandomAffine, "Rotate")
Brightness = special(transforms.ColorJitter, "Brightness")
Contrast = special(transforms.ColorJitter, "Contrast")
Saturation = special(transforms.ColorJitter, "Saturation")
Hue = special(transforms.ColorJitter, "Hue")
Crop = transforms.RandomCrop
class Augmix:
def __init__(self, ops, k=3, alpha=1, beta=1):
self.ops = ops
self.k = k
if isinstance(alpha, Iterable):
self.alpha = alpha
else:
self.alpha = [alpha] * k
self.beta = beta
def get_params(self):
op1, op2, op3 = random.sample(self.ops, 3)
op12 = transforms.Compose([op1, op2])
op123 = transforms.Compose([op1, op2, op3])
return random.sample([op1, op12, op123], 1)[0]
def __call__(self, img):
weights = scipy.stats.dirichlet.rvs(self.alpha)[0]
img_tensor = F.to_tensor(img)
xaug = torch.zeros_like(img_tensor)
for i in range(self.k):
opschain = self.get_params()
temp = weights[i] * F.to_tensor(opschain(img))
xaug += temp
m = scipy.stats.beta.rvs(self.beta, self.beta, size=1)[0]
new_img = m * img_tensor + (1 - m) * xaug
return F.to_pil_image(new_img)
最新文章
- iOS- -安装cocopods
- 尽量多的以 const/enum/inline 替代 #define
- [kuangbin带你飞]专题二十 斜率DP
- 【C++ in Qt5】一个简单的通讯录程序，支持文件存取
- 鸟哥Linux学习笔记03
- SQL-59 按照salary的累计和running_total，其中running_total为前两个员工的salary累计和，其他以此类推。
- 《剑指offer》 树的子结构
- 【转】escape,encodeURI,encodeURIComponent有什么区别?
- 【转】ssh服务器启动和客户端常用操作
- Struts2中 radio标签的详细使用方法
- 深入了解mitmproxy（二）
- C# int转string 每三位加一个逗号
- 读/写锁的实现和应用（高并发状态下的map实现）
- 【Unity笔记】给UGUI元素添加交互性
- bzoj 2818 gcd 线性欧拉函数
- SQL Server 2008内存及I/O性能监控
- 判断序列是否为二叉排序树的后序遍历 python
- java/eclipse/myeclipse建立.properties文件的方法
- [新手教程]windows 2003 php环境搭建详细教程(转)
- Java访问控制权限