On final year of uni, they offered three selection subject to choose: Image Processing, AR/VR, and Digital Forensic. I was trying so hard to be different and chose Digital Forensic, my own worst decision. It was so hard and without any interest on the field, i didn’t actually understand (surviving CS 101). Anyway, a lesson for me (and probably you): just be realistic.

在uni的最后一年，他们提供了三个选择主题供您选择：图像处理，AR / VR和Digital Forensic。 我一直在努力与众不同，因此选择了Digital Forensic，这是我自己最糟糕的决定。 这是如此的艰难，并且对该领域没有任何兴趣，我实际上并没有理解(在CS 101中幸存下来)。 无论如何，对我(可能还有您)来说，这是一个教训：要现实一点。

So because I was so silly to choose another subject instead of image processing, I tried to learn it myself. Turned out, it’s even more fun to learn by yourself. I will list every source to understand Image Processing the easy way on every point. Anyway, Kaggle provides this free course explaining very detailed about Image Processing with Deep Learning.

因此，由于我很愚蠢地选择另一个主题而不是图像处理，所以我尝试自己学习。 事实证明，自己学习更有趣。 我将列出所有资料，以使您轻松地了解图像处理的所有方面。 无论如何，Kaggle都会提供此免费课程，其中详细说明了有关深度学习的图像处理。

In this article, I used Kaggle dataset with X-ray images that classify COVID-19, Viral pneumonia and Normal Chest. There are 219 COVID-19 Positive images, 1341 Normal images and 1345 Viral Pneumonia images.

在本文中，我将Kaggle数据集与X射线图像一起使用，以对COVID-19，病毒性肺炎和正常胸部进行分类。 有219个COVID-19阳性图像，1341正常图像和1345病毒性肺炎图像。

As always, head to my Github for the code❤️

和往常一样，前往我的Github获取代码❤️

I

一世

影像分类(Image Classification)

Image Classification is a method to extract information of image classes digitally. Shortly, it used to classify images. Image Classification can be supervised (you give label to each images) and also unsupervised (the training model learned to classify the images based on the patterns).

图像分类是一种以数字方式提取图像类别信息的方法。 不久，它用于对图像进行分类。 可以监督图像分类(为每个图像添加标签)，也可以不受监督(训练模型可以学习根据模式对图像进行分类)。

This article used Supervised Image Classification with three classes.

本文将监督图像分类与三个类一起使用。

有线电视新闻网 (CNN)

As I mentioned before, I am using Resnet50 layers, which is one of CNN architectures. To know more about CNN, this article explained CNN so clear. To sum it up, the flow of CNN are:

如前所述，我正在使用Resnet50层，这是CNN架构之一。 要了解有关CNN的更多信息，本文对CNN的解释非常清楚。 综上所述，CNN的流为：

1. Convolution Layer (extract feature with filtering)卷积层(通过过滤提取特征)
2. Strides (shifting pixels over the input matrix)步幅(在输入矩阵上移动像素)
3. Padding填充
4. Rectified Linear Unit (RelU) (introduce non-linearity to the network)整流线性单位(RelU)(引入网络非线性)
5. Padding Layer (reduce number of parameters)填充层(减少参数数量)
6. Fully Connected Layer (flatten matrix into vector and feed it to a fully connected neural network layer.)完全连接层(将矩阵展平为向量，并将其馈送到完全连接的神经网络层。)

There are few architecture of CNN (some of the most common is ResNet, VGGNet), but in this post, I will use the ResNet50.

CNN的体系结构很少(最常见的是ResNet，VGGNet)，但是在本文中，我将使用ResNet50。

ResNet or Residual Network uses the residual learning instead of trying to learn some features. Residual can be simply understood as subtraction of feature learned from input of that layer. There are some variants of ResNet other than ResNet50. The core idea of ResNet is introducing a shortcut connection that skips one or more layers.

ResNet或残差网络使用残差学习而不是尝试学习某些功能。 残差可以简单地理解为减去从该层的输入中学到的特征。 除ResNet50之外，还有一些ResNet的变体。 ResNet的核心思想是引入一种跳过一层或多层的快捷连接。

ResNet50 has 50 layers deep, below is the architecture of ResNet50 with 34 layer residual. Here is the paper of the network and here is why ResNet is a good CNN architecture to be used.

ResNet50有50层深，下面是具有34层残留的ResNet50体系结构。 这是网络的文章，这也是ResNet是要使用的良好CNN架构的原因。

数据 (Data)

I actually wanted to try Image Processing for my research but in that time there were so many people said that image training consumed too much time and memory, plus if you don’t have decent laptop/computer, it might be kind of impossible to do so. So I didn’t. But turned out, you could just use Google Colab. I don’t know how much time if I train it on my computer but with Colab it’s not that long like my friends had said.

我实际上是想尝试图像处理来进行研究，但是那时有很多人说图像训练会浪费大量的时间和内存，如果您没有像样的笔记本电脑，那可能是不可能的事情所以。 所以我没有。 但是事实证明，您只能使用Google Colab。 我不知道要在计算机上培训多少时间，但是使用Colab的时间不长，就像我的朋友所说的那样。

If you wanted to access the data from Kaggle, you could just download the JSON API of your Kaggle account, upload to your Colab files then download and unzip the image data to your Colab environment.

如果您想从Kaggle访问数据，则可以下载Kaggle帐户的JSON API，上传到Colab文件，然后下载图像数据并将其解压缩到Colab环境。

! pip install -q kaggle! mkdir ~/.kaggle! cp kaggle.json ~/.kaggle/! chmod 600 ~/.kaggle/kaggle.json! kaggle datasets download -d tawsifurrahman/covid19-radiography-database! unzip "covid19-radiography-database.zip"

You can also access data from your Google Drive with mounting your Colab session and access your image data location from the Drive.

您还可以通过挂载Colab会话从Google云端硬盘访问数据，并从云端硬盘访问图像数据位置。

from google.colab import drivedrive.mount('/content/drive')

If you want to access the data from an open source website, you can just use request library from Python. To take an example, here’s how to download 20GB data from Physio Net.

如果要从开源网站访问数据，则可以仅使用Python中的请求库。 举个例子，这里是如何从Physio Net下载20GB数据的方法。

import requests, zipfile, iozip_file_url = "https://physionet.org/static/published-projects/siena-scalp-eeg/siena-scalp-eeg-database-1.0.0.zip"r = requests.get(zip_file_url)z = zipfile.ZipFile(io.BytesIO(r.content))z.extractall()

预处理 (Pre-Processing)

Library

图书馆

Some libraries that I used

我使用的一些库

from fastai.vision import *from fastai.metrics import error_rateimport torchimport torch.nn as nnimport torch.optim as optimfrom torch.optim import lr_schedulerfrom torch.autograd import Variableimport numpy as npimport timeimport osfrom fastai.callbacks import ActivationStats%matplotlib inlineimport matplotlib.pyplot as pltplt.style.use('seaborn-whitegrid')

Access the image from folder

从文件夹访问图像

Define the folder path

定义文件夹路径

path = Path('/content/COVID-19 Radiography Database/')folder = pathfolder.mkdir(parents=True, exist_ok=True)

Extract the image data bunch

提取图像数据串

np.random.seed(42)data = ImageDataBunch.from_folder(path, train='.', valid_pct=0.35, ds_tfms=get_transforms(), size=224, num_workers=4).normalize(imagenet_stats)

使用CNN Resnet50进行培训 (Training with CNN Resnet50)

With fastai, we can instantly use the Resnet model by the function of cnn_learner. Don’t forget to define the accuracy metrics so that we can see the loss and accuracy.

使用fastai，我们可以通过cnn_learner的功能立即使用Resnet模型。 不要忘记定义准确性指标，以便我们可以看到损失和准确性。

To reduce time, I will just use four epochs.

为了减少时间，我将仅使用四个时期。

learn = cnn_learner(data, models.resnet50, metrics=accuracy)learn.fit_one_cycle(4)

With the model, we can find the optimal learning rate to get the better performance. Unfreezing the model is used to let the model continue the training. In this case, unfreezing will let the model to train again with the better learning rate after we got the observation.

通过该模型，我们可以找到最佳学习率以获得更好的性能。 解冻模型用于使模型继续训练。 在这种情况下，解冻将使模型在得到观察后以更好的学习率再次训练。

learn.unfreeze()learn.lr_find()learn.recorder.plot(suggestion=True)

Next training with better learning rate can train for just 2 epochs.

具有较高学习率的下一次训练只能训练2个时期。

结果 (Result)

The very good thing that I really love using FastAI is that the model evaluation is so so easy. We can just interpret the classification model like the loss prediction, confusion matrix, most confused.

我真正喜欢使用FastAI的好处是，模型评估是如此简单。 我们可以解释分类模型，例如损失预测，混淆矩阵，最混淆的模型。

interp = ClassificationInterpretation.from_learner(learn)losses, idxs = interp.top_losses()

First, we can plot the loss of prediction from the model.

首先，我们可以绘制模型的预测损失。

interp.plot_top_losses(9, figsize=(15,11))

And then the Confusion Matrix.

然后是混淆矩阵。

interp.plot_confusion_matrix(figsize=(12,12), dpi=100)

结论 (Conclusion)

It’s very convenient to apply many AI methods now, especially with FastAI Image Processing. I think the most important thing to be on studying Machine Learning is to be creative, innovative and never stops learning new things!

现在应用许多AI方法非常方便，尤其是使用FastAI图像处理时。 我认为学习机器学习最重要的是要有创造力，创新能力，并且要不断学习新事物！

Thank you for reading!

感谢您的阅读！

更多学习资料 (More Source to Learn)