【图像处理】图像常见读取方式速度对比

0.简介

即使随着现在得发展，CPU的处理速度提高和SSD的出现提高了数据的读取速度，但是对于超大规模的训练过程（ImageNet），读取图像的时间依然是一个较大的开销，因此本文旨在对常见的框架技术的图像读取方式进行对比。

图像属性：（427，640，3）

附上我的学习链接：https://zhuanlan.zhihu.com/p/30383580

1.对比过程

1.1 OpenCV

import cv2
import time

image = '1.jpg'
MAXN = 1000

time1 = time.time()
for i in range(MAXN):
    img = cv2.imread(image)
    # img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
print(MAXN / (time.time() - time1))

449.38982663317097 imgs / second

import cv2
import time

image = '1.jpg'
MAXN = 1000

time1 = time.time()
for i in range(MAXN):
    img = cv2.imread(image)
    img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
print(MAXN / (time.time() - time1))

346.7856687525765 imgs / second

1.2 scipy

from scipy.misc import imread
import time

image = '1.jpg'
MAXN = 1000

time1 = time.time()
for i in range(MAXN):
    img = imread(image)
print(MAXN / (time.time() - time1))

322.0913625719489 imgs / second

1.3 skimage

from skimage import io
import time

image = '1.jpg'
MAXN = 1000

time1 = time.time()
for i in range(MAXN):
    img = io.imread(image)
print(MAXN / (time.time() - time1))

328.2046402456301 imgs / second

1.4 PIL

from PIL import Image
import time
import numpy as np

image = '1.jpg'
MAXN = 1000

time1 = time.time()
for i in range(MAXN):
    img = Image.open(image)
    img = np.array(img)
print(MAXN / (time.time() - time1))

281.86613847589075 imgs / second

from PIL import Image
import time
import numpy as np

image = '1.jpg'
MAXN = 1000

time1 = time.time()
for i in range(MAXN):
    img = Image.open(image)
    # img = np.array(img)
print(MAXN / (time.time() - time1))

17242.91258304282 imgs / second

注：大部分时间都消耗在了numpy转换上了。

1.5 MXNet（存在问题！）

import mxnet as mx
import time

image = '1.jpg'
MAXN = 1000

time1 = time.time()
for i in range(MAXN):
    img = mx.image.imdecode(open(image, 'rb').read())
mx.nd.waitall()
print(MAXN / (time.time() - time1))

318.6462058057628 imgs / second

注：不知道为啥，速度感觉是错误的。

1.6 Tensorflow

import tensorflow as tf
import time

image = '1.jpg'
MAXN = 1000

time1 = time.time()
for i in range(MAXN):
    img = tf.gfile.FastGFile(image, 'rb').read()
    img = tf.image.decode_jpeg(img)
print(MAXN / (time.time() - time1))

1634.8531350343824 imgs / second

注：这个方法的速度还是很快的，一开始的时候直接用tf.gfile.FastGFile(image)，显示编码错误，改成'rb'方式就好了。

此处附上一位大神的讲解（关于Tensorflow数据读取方式）：
https://zhuanlan.zhihu.com/p/27238630