快捷導(dǎo)航

詳解python opencv、scikit-image和PIL圖像處理庫比較

更新時(shí)間：2019年12月26日 14:25:38 作者：Oldpan

這篇文章主要介紹了詳解python opencv、scikit-image和PIL圖像處理庫比較，文中通過示例代碼介紹的非常詳細(xì)，對(duì)大家的學(xué)習(xí)或者工作具有一定的參考學(xué)習(xí)價(jià)值，需要的朋友們下面隨著小編來一起學(xué)習(xí)學(xué)習(xí)吧

進(jìn)行深度學(xué)習(xí)時(shí)，對(duì)圖像進(jìn)行預(yù)處理的過程是非常重要的，使用pytorch或者TensorFlow時(shí)需要對(duì)圖像進(jìn)行預(yù)處理以及展示來觀看處理效果，因此對(duì)python中的圖像處理框架進(jìn)行圖像的讀取和基本變換的掌握是必要的，接下來python中幾個(gè)基本的圖像處理庫進(jìn)行縱向?qū)Ρ取?/p>

項(xiàng)目地址：https://github.com/Oldpan/Pytorch-Learn/tree/master/Image-Processing

比較的圖像處理框架：

PIL
scikit-image
opencv-python

PIL：

由于PIL僅支持到Python 2.7，加上年久失修，于是一群志愿者在PIL的基礎(chǔ)上創(chuàng)建了兼容的版本，名字叫Pillow，支持最新Python 3.x，又加入了許多新特性，因此，我們可以直接安裝使用Pillow。

摘自廖雪峰的官方網(wǎng)站

scikit-image

scikit-image is a collection of algorithms for image processing. It is available free of charge and free of restriction. We pride ourselves on high-quality, peer-reviewed code, written by an active community of volunteers.
摘自官網(wǎng)的介紹，scikit-image的更新還是比較頻繁的，代碼質(zhì)量也很好。

opencv-python

opencv的大名就不要多說了，這個(gè)是opencv的python版

# Compare Image-Processing Modules
# Use Transforms Module of torchvision
#        &&&
# 對(duì)比python中不同的圖像處理模塊
# 并且使用torchvision中的transforms模塊進(jìn)行圖像處理

# packages
from PIL import Image
from skimage import io, transform
import cv2

import torchvision.transforms as transforms
import matplotlib.pyplot as plt
%matplotlib inline

img_PIL = Image.open('./images/dancing.jpg')
img_skimage = io.imread('./images/dancing.jpg')
img_opencv = cv2.imread('./images/dancing.jpg')
img_plt = plt.imread('./images/dancing.jpg')

loader = transforms.Compose([
  transforms.ToTensor()]) # 轉(zhuǎn)換為torch.tensor格式


print('The shape of \n img_skimage is {}\n img_opencv is {}\n img_plt is {}\n'.format(img_skimage.shape, img_opencv.shape, img_plt.shape))
print('The type of \n img_skimage is {}\n img_opencv is {}\n img_plt is {}\n'.format(type(img_skimage), type(img_opencv), type(img_plt)))

The shape of
img_skimage is (444, 444, 3)
img_opencv is (444, 444, 3)
img_plt is (444, 444, 3)
The size of
img_PIL is (444, 444)
The mode of
img_PIL is RGB
The type of
img_skimage is <class 'numpy.ndarray'>
img_opencv is <class 'numpy.ndarray'>
img_plt is <class 'numpy.ndarray'>
img_PIL if <class 'PIL.JpegImagePlugin.JpegImageFile'>

# 定義一個(gè)圖像顯示函數(shù)
def my_imshow(image, title=None):
  plt.imshow(image)
  if title is not None:
    plt.title(title)
  plt.pause(0.001) # 這里延時(shí)一下，否則圖像無法加載

plt.figure()
my_imshow(img_skimage, title='img_skimage')
# 可以看到opencv讀取的圖像打印出來的顏色明顯與其他不同
plt.figure()
my_imshow(img_opencv, title='img_opencv')
plt.figure()
my_imshow(img_plt, title='img_plt')

# opencv讀出的圖像顏色通道為BGR，需要對(duì)此進(jìn)行轉(zhuǎn)換
img_opencv = cv2.cvtColor(img_opencv, cv2.COLOR_BGR2RGB)
plt.figure()
my_imshow(img_opencv, title='img_opencv_new')

toTensor = transforms.Compose([transforms.ToTensor()])

# 尺寸變化、縮放
transform_scale = transforms.Compose([transforms.Scale(128)])
temp = transform_scale(img_PIL)
plt.figure()
my_imshow(temp, title='after_scale')

# 隨機(jī)裁剪
transform_randomCrop = transforms.Compose([transforms.RandomCrop(32, padding=4)])
temp = transform_scale(img_PIL)
plt.figure()
my_imshow(temp, title='after_randomcrop')

# 隨機(jī)進(jìn)行水平翻轉(zhuǎn)（0.5幾率）
transform_ranHorFlip = transforms.Compose([transforms.RandomHorizontalFlip()])
temp = transform_scale(img_PIL)
plt.figure()
my_imshow(temp, title='after_ranhorflip')

# 隨機(jī)裁剪到特定大小
transform_ranSizeCrop = transforms.Compose([transforms.RandomSizedCrop(128)])
temp = transform_ranSizeCrop(img_PIL)
plt.figure()
my_imshow(temp, title='after_ranSizeCrop')

# 中心裁剪
transform_centerCrop = transforms.Compose([transforms.CenterCrop(128)])
temp = transform_centerCrop(img_PIL)
plt.figure()
my_imshow(temp, title='after_centerCrop')

# 空白填充
transform_pad = transforms.Compose([transforms.Pad(4)])
temp = transform_pad(img_PIL)
plt.figure()
my_imshow(temp, title='after_padding')

# 標(biāo)準(zhǔn)化是在整個(gè)數(shù)據(jù)集中對(duì)所有圖像進(jìn)行取平均和均方差，演示圖像數(shù)量過少無法進(jìn)行此操作
# print(train_data.mean(axis=(0,1,2))/255)
# print(train_data.std(axis=(0,1,2))/255)
# transform_normal = transforms.Compose([transforms.Normalize()])

# Lamdba使用用戶自定義函數(shù)來對(duì)圖像進(jìn)行剪裁
# transform_pad = transforms.Compose([transforms.Lambda()])