python+opencv 图像预处理

from PIL import Image,ImageDraw,ImageFont
import numpy as np
import random
import string
import cv2
# 生成随机字符串
for i in range(1,10000):
    strnum = random.randint(2,10)
    ran_str = "".join(random.sample(string.ascii_letters + string.digits, strnum))
    font = ImageFont.truetype(‘D:\Multimedia\RosewoodStd-Regular.otf‘, 60)  # otf和ttf 是都可以使用。
    image = Image.new("RGB", (300, 200), (255, 255, 255)) 
    draw = ImageDraw.Draw(image)
    if strnum <=5:
        #  参数一 写入位置 参数二 文本 参数三 字体格式 fill 为文本颜色 默认为白色
        draw.text((80, 80), ran_str, font=font, fill=‘Black‘)
    else:
        draw.text((20, 80), ran_str, font=font, fill=‘Black‘)
    path = "D:\AdobeVFR_release\\rosewood"+"\\rosewood_regular_"+str(i)+".jpeg"
    image.save(path)
    if(i%10==0):
        print("has already save %d images"%i)

import  cv2
from skimage import data_dir,io,color
import numpy as np
import random
sigma = random.uniform(2.5,3.5)
# 高斯噪声函数，这里写的是按像素点单个处理，可以建一个高斯随机数的矩阵
def GaussianNoise (img ,means =0,sigma =1.5):  
    r = img[:,:,0].flatten()
    g = img[:,:,1].flatten()
    b = img[:,:,2].flatten()
    for i in range(img.shape[0]*img.shape[1]):
        r[i] = r[i]+random.gauss(0,sigma)
        g[i] = g[i]+random.gauss(0,sigma)
        b[i] = b[i]+random.gauss(0,sigma)
    img[:,:,0] = r.reshape([img.shape[0],img.shape[1]])
    img[:,:,1] = r.reshape([img.shape[0],img.shape[1]])
    img[:,:,2] = r.reshape([img.shape[0],img.shape[1]])
    return img
# 几何变化函数，主要功能是扭曲
def Geometric_changes(image,width,height):
    pts1 = np.float32([[50, 50], [200, 50], [50, 200]])
    x = random.randint(50,100)
    y = random.randint(200,250)
    z = random.randint(10,50)
    pts2 = np.float32([[z, x], [200, 50], [x, y]])
    M = cv2.getAffineTransform(pts1, pts2)
    image_0 = cv2.warpAffine(image, M, (width, height))
    return image_0
path =‘D:\AdobeVFR_release\sythetic‘
string = path+‘/*.jpeg‘
”“”
io.ImageCollrction 将图片路径整理成一个list
”“”
coll = io.ImageCollection(string) 
a = np.array(coll)

for i in range(95,len(a)):
    height = a[i].shape[0]
    width = a[i].shape[1]
    image_0 = GaussianNoise(a[i])
    image_0 = cv2.GaussianBlur(image_0,(5,5),sigma)  # 高斯模糊
    image_0 = Geometric_changes(image_0, width, height)

    # 缩放函数，fx，fy为缩放因子 interpolation有五种，

    #INTER_AREA　基于局部像素的重采样 图像缩小时候，该方法可以避免波纹
    #INTER_NEAREST 最近邻插值法      适合放大
    #INTER_LINEAR 双线性插值法       默认
    #INTER_CUBIC  基于4x4像素邻域的3次插值法  适合放大
    #INTER_LANCZOS4 - 基于8x8像素邻域的Lanczos插值

    image_0 = cv2.resize(image_0, None, fx=random.uniform(5 / 6, 7 / 6), fy=1, interpolation=cv2.INTER_AREA)
    
    path = "D:\AdobeVFR_release\sydata"+"\d"+str(i)+".jpeg"
    cv2.imwrite(path, image_0)
    if(i%10==0):
        print("has already %d image" % i)