cnn对网络数据预处理_CNN中的数据预处理和网络构建

cnn对网络数据预处理

In this article, we will go through the end-to-end pipeline of training convolution neural networks, i.e. organizing the data into directories, preprocessing, data augmentation, model building, etc.

在本文中，我们将遍历训练卷积神经网络的端到端管道，即将数据组织到目录，预处理，数据扩充，模型构建等中。

We will spend a good amount of time on data preprocessing techniques commonly used with image processing. This is because preprocessing takes about 50–80% of your time in most deep learning projects, and knowing some useful tricks will help you a lot in your projects. We will be using the flowers dataset from Kaggle to demonstrate the key concepts. To get into the codes directly, an accompanying notebook is published on Kaggle(Please use a CPU for running the initial parts of the code and GPU for model training).

我们将在图像处理常用的数据预处理技术上花费大量时间。 这是因为在大多数深度学习项目中，预处理需要花费您大约50-80％的时间，并且了解一些有用的技巧将对您的项目有很大帮助。 我们将使用来自Kaggle的flowers数据集来演示关键概念。 为了直接进入代码，在Kaggle上发布了一个附带的笔记本 (请使用CPU运行代码的初始部分，并使用GPU进行模型训练)。

导入数据集 (Importing the dataset)

Let’s begin with importing the necessary libraries and loading the dataset. This is a requisite step in every data analysis process.

让我们从导入必要的库并加载数据集开始。 这是每个数据分析过程中的必要步骤。

# Importing necessary librariesimport keras
import tensorflow
from skimage import io
import os
import glob
import numpy as np
import random
import matplotlib.pyplot as plt
%matplotlib inline# Importing and Loading the data into data frame
#class 1 - Rose, class 0- DaisyDATASET_PATH = '../input/flowers-recognition/flowers/'
flowers_cls = ['daisy', 'rose']# glob through the directory (returns a list of all file paths)flower_path = os.path.join(DATASET_PATH, flowers_cls[1], '*')
flower_path = glob.glob(flower_path)# access some element (a file) from the list
image = io.imread(flower_path[251])

数据预处理 (Data Preprocessing)

图片-频道和大小 (Images — Channels and Sizes)

Images come in different shapes and sizes. They also come through different sources. For example, some images are what we call “natural images”, which means they are taken in color, in the real world. For example:

图像具有不同的形状和大小。 它们也来自不同的来源 。 例如，有些图像被我们称为“自然图像”，这意味着它们是在现实世界中以彩色拍摄的。 例如：

• A picture of a flower is a natural image.
  花的图片是自然图像。

• An X-ray image is not a natural image.

  X射线图像不是自然图像。

Taking all these variations into consideration, we need to perform some pre-processing on any image data. RGB is the most popular encoding format, and most “natural images” we encounter are in RGB. Also, among the first step of data pre-processing is to make the images of the same size. Let’s move on to how we can change the shape and form of images.

考虑到所有这些变化，我们需要对任何图像数据进行一些预处理。 RGB是最流行的编码格式，我们遇到的大多数“自然图像”都是RGB。 同样，数据预处理的第一步是使图像大小相同。 让我们继续介绍如何更改图像的形状和形式。

# plotting the original image and the RGB channels
f, (ax1, ax2, ax3, ax4) = plt.subplots(1, 4, sharey=True)
f.set_figwidth(15)
ax1.imshow(image)# RGB channels
# CHANNELID : 0 for Red, 1 for Green, 2 for Blue. 
ax2.imshow(image[:, : , 0]) #Red
ax3.imshow(image[:, : , 1]) #Green
ax4.imshow(image[:, : , 2]) #Blue
f.suptitle('Different Channels of Image')

形态转换 (Morphological Transformations)

The term morphological transformation refers to any modification involving the shape and form of the images. These are very often used in image analysis tasks. Although they are used with all types of images, they are especially powerful for images that are not natural (come from a source other than a picture of the real world). The typical transformations are erosion, dilation, opening, and closing. Let’s now look at some code to implement these morphological transformations.

术语形态变换是指涉及图像形状和形式的任何修饰。 这些通常在图像分析任务中使用。 尽管它们可用于所有类型的图像，但对于非自然的图像(来自真实世界的图片以外的其他来源)，它们尤其强大。 典型的转换是侵蚀，膨胀，打开和关闭。 现在让我们看一些实现这些形态转换的代码。

1.Thresholding

1.阈值

One of the simpler operations where we take all the pixels whose intensities are above a certain threshold and convert them to ones; the pixels having value less than the threshold are converted to zero. This results in a binary image.

一种比较简单的操作，其中我们将强度高于特定阈值的所有像素都转换为像素； 值小于阈值的像素将转换为零。 这将产生二进制图像 。

# bin_image will be a (240, 320) True/False array
#The range of pixel varies between 0 to 255
#The pixel having black is more close to 0 and pixel which is white is more close to 255
# 125 is Arbitrary heuristic measure halfway between 1 and 255 (the range of image pixel) 
bin_image = image[:, :, 0] > 125
plot_image([image, bin_image], cmap='gray')

2.Erosion, Dilation, Opening & Closing

2.侵蚀，膨胀，开合

Erosion shrinks bright regions and enlarges dark regions. Dilation on the other hand is exact opposite side — it shrinks dark regions and enlarges the bright regions.

侵蚀使明亮的区域收缩，使黑暗的区域扩大。 另一方面， 膨胀恰好在相反的一面-它会缩小深色区域并扩大明亮区域。

Opening is erosion followed by dilation. Opening can remove small bright spots (i.e. “salt”) and connect small dark cracks. This tends to “open” up (dark) gaps between (bright) features.

开放是侵蚀，然后是膨胀。 开口可以去除小的亮点(即“盐”)并连接小的暗裂纹。 这往往会“打开”(亮)特征之间的(暗)间隙。

Closing is dilation followed by erosion. Closing can remove small dark spots (i.e. “pepper”) and connect small bright cracks. This tends to “close” up (dark) gaps between (bright) features.

关闭是扩张，然后是侵蚀。 关闭可以消除小黑点(即“胡椒”)并连接小亮点。 这往往会“缩小”(亮)特征之间的(暗)间隙。

All these can be done using the skimage.morphology module. The basic idea is to have a circular disk of a certain size (3 below) move around the image and apply these transformations using it.

所有这些都可以使用skimage.morphology模块完成。 基本思想是让一定大小(以下为3)的圆盘在图像周围移动并使用该图像应用这些转换。

from skimage.morphology import binary_closing, binary_dilation, binary_erosion, binary_opening
from skimage.morphology import selem# use a disk of radius 3
selem = selem.disk(3)# oprning and closing
open_img = binary_opening(bin_image, selem)
close_img = binary_closing(bin_image, selem)
# erosion and dilation
eroded_img = binary_erosion(bin_image, selem)
dilated_img = binary_dilation(bin_image, selem)
plot_image([bin_image, open_img, close_img, eroded_img, dilated_img], cmap='gray')

正常化 (Normalisation)

Normalisation is the most crucial step in the pre-processing part. This refers to rescaling the pixel values so that they lie within a confined range. One of the reasons to do this is to help with the issue of propagating gradients. There are multiple ways to normalize images that we will be talking about.

标准化是预处理部分中最关键的步骤。 这是指重新缩放像素值，以使其位于限定范围内。 这样做的原因之一是帮助解决传播梯度的问题。 我们将讨论多种标准化图像的方法。

#way1-this is common technique followed in case of RGB images 
norm1_image = image/255#way2-in case of medical Images/non natural images 
norm2_image = image - np.min(image)/np.max(image) - np.min(image)#way3-in case of medical Images/non natural images 
norm3_image = image - np.percentile(image,5)/ np.percentile(image,95) - np.percentile(image,5)
plot_image([image, norm1_image, norm2_image, norm3_image], cmap='gray')

增广 (Augmentation