目标检测：python实现多种图像数据增强的方法（光照，对比度，遮挡，模糊）
Posted 告白少年
tags:
篇首语：本文由小常识网(cha138.com)小编为大家整理，主要介绍了目标检测：python实现多种图像数据增强的方法（光照，对比度，遮挡，模糊）相关的知识，希望对你有一定的参考价值。
图像数据增强的内容（可根据需要自定义选择）：
1.直方图均衡化
2.clahe自适应对比度直方图均衡化
3.白平衡
4.亮度增强
5.亮度，饱和度，对比度增强
6.去除图像上的高光部分
7.自适应亮度增强
8.随机遮挡
9.图像高斯模糊
10.压缩图像
# -*- coding: utf-8 -*-
"""
******不改变原始xml的一些数据增强方法 type 1-10*******
把增强后的图像和xml一起放入新的文件夹
rootpath:picture_xml原始路径
savepath：picture_xml保存路径
*******改变原始 xml的一些数据增强方法 type 11-15******
修改图片的同时修改对应的xml
file_path:传入类别的信息txt，最好和生成labelmap的顺序一致
rootpath:picture_xml原始路径
savepath：picture_xml保存路径
11:自定义裁剪，图像大小 w,h，例如 w=400,h=600
12：自定义平移，平移比例 w,h [0-1] 例如w=0.1,h=0,2
13：自定义缩放，调整图像大小 w,h,例如 w=400,h=600
14：图像翻转
15:图像任意旋转，传入旋转角度列表anglelist=[90,-90]
"""
import cv2
import random
import math
import os,shutil
import numpy as np
from PIL import Image, ImageStat
from skimage import exposure
import matplotlib.pyplot as plt
import tensorlayer as tl
from lxml.etree import Element, SubElement, tostring
import xml.etree.ElementTree as ET
def hisColor_Img(path):
"""
对图像直方图均衡化
:param path: 图片路径
:return: 直方图均衡化后的图像
"""
img = cv2.imread(path)
ycrcb = cv2.cvtColor(img, cv2.COLOR_BGR2YCR_CB)
channels = cv2.split(ycrcb)
cv2.equalizeHist(channels[0], channels[0]) #equalizeHist(in,out)
cv2.merge(channels, ycrcb)
img_eq=cv2.cvtColor(ycrcb, cv2.COLOR_YCR_CB2BGR)
return img_eq
def clahe_Img(path,ksize):
"""
:param path: 图像路径
:param ksize: 用于直方图均衡化的网格大小，默认为8
:return: clahe之后的图像
"""
image = cv2.imread(path, cv2.IMREAD_COLOR)
b, g, r = cv2.split(image)
clahe = cv2.createCLAHE(clipLimit=2.0, tileGridSize=(ksize,ksize))
b = clahe.apply(b)
g = clahe.apply(g)
r = clahe.apply(r)
image = cv2.merge([b, g, r])
return image
def whiteBalance_Img(path):
"""
对图像白平衡处理
"""
img = cv2.imread(path)
b, g, r = cv2.split(img)
Y = 0.299 * r + 0.587 * g + 0.114 * b
Cr = 0.5 * r - 0.419 * g - 0.081 * b
Cb = -0.169 * r - 0.331 * g + 0.5 * b
Mr = np.mean(Cr)
Mb = np.mean(Cb)
Dr = np.var(Cr)
Db = np.var(Cb)
temp_arry = (np.abs(Cb - (Mb + Db * np.sign(Mb))) < 1.5 * Db) & (
np.abs(Cr - (1.5 * Mr + Dr * np.sign(Mr))) < 1.5 * Dr)
RL = Y * temp_arry
# 选取候选白点数的最亮10%确定为最终白点，并选择其前10%中的最小亮度值
L_list = list(np.reshape(RL, (RL.shape[0] * RL.shape[1],)).astype(np.int))
hist_list = np.zeros(256)
min_val = 0
sum = 0
for val in L_list:
hist_list[val] += 1
for l_val in range(255, 0, -1):
sum += hist_list[l_val]
if sum >= len(L_list) * 0.1:
min_val = l_val
break
# 取最亮的前10%为最终的白点
white_index = RL < min_val
RL[white_index] = 0
# 计算选取为白点的每个通道的增益
b[white_index] = 0
g[white_index] = 0
r[white_index] = 0
Y_max = np.max(RL)
b_gain = Y_max / (np.sum(b) / np.sum(b > 0))
g_gain = Y_max / (np.sum(g) / np.sum(g > 0))
r_gain = Y_max / (np.sum(r) / np.sum(r > 0))
b, g, r = cv2.split(img)
b = b * b_gain
g = g * g_gain
r = r * r_gain
# 溢出处理
b[b > 255] = 255
g[g > 255] = 255
r[r > 255] = 255
res_img = cv2.merge((b, g, r))
return res_img
def bright_Img(path,ga,flag):
"""
亮度增强 Tensorlayer
:param ga: ga为gamma值，>1亮度变暗，<1亮度变亮
:param flag:True: 亮度值为(1-ga,1+ga)
False:亮度值为ga,默认为1
:return: 亮度增强后的图像
"""
image = tl.vis.read_image(path)
tenl_img = tl.prepro.brightness(image, gamma=ga, is_random=flag)
return tenl_img
def illumination_Img(path,ga,co,sa,flag):
"""
亮度,饱和度，对比度增强 Tensorlayer
:param ga: ga为gamma值，>1亮度变暗，<1亮度变亮
:param co: 对比度值，1为原始值
:param sa: 饱和度值，1为原始值
:param flag:True: 亮度值为(1-ga,1+ga)，对比度(1-co,1+co)，饱和度(1-sa,1+sa)
False:亮度值为ga,对比度co,饱和度sa
:return:增强后的结果
"""
image = tl.vis.read_image(path)
tenl_img= tl.prepro.illumination(image, gamma=ga, contrast=co , saturation=sa, is_random=flag)
return tenl_img
def create_mask(imgpath):
image = cv2.imread(imgpath, cv2.IMREAD_GRAYSCALE)
_, mask = cv2.threshold(image, 200, 255, cv2.THRESH_BINARY)
return mask
def xiufu_Img(imgpath,maskpath):
"""
去除图像上的高光部分
"""
src_ = cv2.imread(imgpath)
mask = cv2.imread(maskpath, cv2.IMREAD_GRAYSCALE)
#缩放因子(fx,fy)
res_ = cv2.resize(src_,None,fx=0.6, fy=0.6, interpolation = cv2.INTER_CUBIC)
mask = cv2.resize(mask,None,fx=0.6, fy=0.6, interpolation = cv2.INTER_CUBIC)
dst = cv2.inpaint(res_, mask, 10, cv2.INPAINT_TELEA)
return dst
def image_brightness(rgb_image):
'''
检测图像亮度(基于RMS)
'''
stat = ImageStat.Stat(rgb_image)
r, g, b = stat.rms
return math.sqrt(0.241*(r**2) + 0.691*(g**2) + 0.068*(b**2))
def calc_gamma(brightness):
return brightness/127.0
def image_gamma_transform(pil_im, gamma):
image_arr = np.array(pil_im)
image_arr2 = exposure.adjust_gamma(image_arr, gamma)
if len(image_arr.shape) == 3: # 格式为(height(rows), weight(colums), 3)
r = Image.fromarray(np.uint8(image_arr[:, :, 0]))
g = Image.fromarray(np.uint8(image_arr[:, :, 1]))
b = Image.fromarray(np.uint8(image_arr[:, :, 2]))
image = Image.merge("RGB", (r, g, b))
return image
elif len(image_arr.shape) == 2: # 格式为(height(rows), weight(colums))
return Image.fromarray(np.uint8(image_arr))
def autobright_Img(rootpath,savepath):
"""
自适应亮度增强
"""
list=os.listdir(rootpath)
for i in range(0, len(list)):
path = os.path.join(rootpath, list[i])
if os.path.isfile(path):
if list[i].endswith("jpg") or list[i].endswith("JPG") or list[i].endswith("png") or list[i].endswith("PNG"):
print("adjust_bright running....")
print(list[i])
im = Image.open(path)
brightness = image_brightness(im)
newimage = np.array(image_gamma_transform(im, calc_gamma(brightness)))
newname="adjust_bright"+list[i][:-4]
saveflie=os.path.join(savepath, newname+".jpg")
plt.imsave(saveflie, newimage)
shutil.copyfile(os.path.join(rootpath, list[i][:-4] + ".xml"),
os.path.join(savepath, newname + ".xml"))
def probability_random_event(rate, event):
"""随机变量的概率函数"""
#
# 参数rate为list<int>
# 返回概率事件的下标索引
start = 0
index = 0
randnum = random.randint(1, sum(rate))
for index, scope in enumerate(rate):
start += scope
if randnum <= start:
break
return event[index]
def erase_Img(root_path,save_path):
"""
随机遮挡
"""
for file in os.listdir(root_path):
file_name, extension = os.path.splitext(file)
if extension == '.xml':
print("erase running....")
print(file_name+".jpg")
xml_path = os.path.join(root_path, file)
tree = ET.parse(xml_path)
root = tree.getroot()
image_path = os.path.join(root_path,file_name+'.jpg')
image = cv2.imread(image_path)
for obj in root.findall('object'):
#文件夹图像遮挡的比例，参数可修改
is_erase = probability_random_event([7,3],[True,False])
if is_erase:
#boundingbox遮挡的方向，参数可修改
erase_orientation = probability_random_event([6,2,1,1],
['down', 'up', 'left','right'])
#遮挡方块的大小，参数可修改
erase_scope = random.uniform(0.1,0.3)
xml_box = obj.find('bndbox')
_xmin = int(xml_box.find('xmin').text)
_xmax = int(xml_box.find('xmax').text)
_ymin = int(xml_box.find('ymin').text)
_ymax = int(xml_box.find('ymax').text)
box_width = _xmax - _xmin
box_height = _ymax - _ymin
new_xmin, new_xmax, new_ymin, new_ymax = _xmin, _xmax, _ymin, _ymax
if erase_orientation == 'down':
new_ymax = int(_ymax - box_height*erase_scope)
image[new_ymax:_ymax,new_xmin:new_xmax,:] = 255
if erase_orientation == 'up':
new_ymin = int(_ymin + box_height*erase_scope)
image[_ymin:new_ymin,new_xmin:new_xmax,:] = 255
if erase_orientation == 'left':
new_xmin = int(_xmin + box_width*erase_scope)
image[new_ymin:new_ymax,_xmin:new_xmin,:] = 255
if erase_orientation == 'right':
new_xmax = int(_xmax - box_width*erase_scope)
image[new_ymin:new_ymax,new_xmax:_xmax,:] = 255
cv2.imwrite(os.path.join(save_path,"earse_"+file_name+'.jpg'),image)
xml_box.find('xmin').text = str(new_xmin)
xml_box.find('ymin').text = str(new_ymin)
xml_box.find('xmax').text = str(new_xmax)
xml_box.find('ymax').text = str(new_ymax)
tree.write(os.path.join(save_path,"earse_"+file_name+'.xml'))
def blur_Img(rootpath,savepath,ksize,new_rate):
"""
随机模糊图像
"""
img_list=[]
for imgfiles in os.listdir(rootpath):
if (imgfiles.endswith("jpg") or imgfiles.endswith("JPG")):
img_list.append(imgfiles)
filenumber = len(img_list)
rate =new_rate # 自定义抽取文件夹中图片的比例，参数可修改
picknumber = int(filenumber * rate)
sample = random.sample(img_list, picknumber)
for name in sample:
print("blur running....")
print(name)
namepath=os.path.join(rootpath,name)
ori_img = cv2.imread(namepath)
size = random.choice(ksize) # 设置高斯核的大小，参数可修改，size>9，小于9图像没有变化
kernel_size = (size, size)
image = cv2.GaussianBlur(ori_img, ksize=kernel_size, sigmaX=0, sigmaY=0)
cv2.imwrite(os.path.join(savepath,"blur_"+name), image)
shutil.copyfile(os.path.join(rootpath, name.split(".")[0] + ".xml"),
os.path.join(savepath, "blur_"+name.split(".")[0] + ".xml"))
def compress_Img(infile_path, outfile_path,pic_size):
"""
压缩图像
不改变图片尺寸压缩到指定大小
:param infile: 压缩源文件
:param outfile: 压缩文件保存地址
"""
count=0
for infile in os.listdir(infile_path):
if infile.endswith(".jpg") or infile.endswith(".GPG") or \\
infile.endswith(".jpeg") or infile.endswith(".GPEG"):
print("compress_ running....")
print(infile)
filename, extend_name = os.path.splitext(infile)
img_path=os.path.join(infile_path,infile)
imgsaved_path=os.path.join(outfile_path,"compress_"+infile)
img = cv2.imread(img_path, 1)
以上是关于目标检测：python实现多种图像数据增强的方法（光照，对比度，遮挡，模糊）的主要内容，如果未能解决你的问题，请参考以下文章
目标检测(10) Mosaic 数据增强方法，附Python完整代码
Python3 实现批量图像数据增强（扩增）并复制xml标签文件目标检测笔记
目标检测数据增强：YOLO官方数据增强实现/imgaug的简单使用