Computer vision: How do you handle image data? - Python教程

- [Instructor] In this chapter, we would continue our discussion of AutoML by exploring its use for image data. Businesses deal with various image data including images of receipts, clients, and product. As an example, an eCommerce website may require a tool to confirm defects through images shared by customers. We can use machine learning techniques in these types of scenarios, and AutoML can help us here. There are some key differences between image and tabular data, which we examined in the first chapter. First, image data is not neatly arranged into rows and columns. Instead, each image is unstructured, and we are interested in the domain it contains. We can use tools like NumPy to read image into three dimensional arrays for processing. The field that caters to the understanding and processing of images is called Computer Vision. And a variety of tasks sit under this field. Image classification is a common one where we want to separate images of one thing from other images. Facial recognition is another example. It involves identification of a customer or a person by facial features. This task has become a key part of KYC for many businesses. There is also object detection, which can be used for security in places like airports and shopping malls. All these are tasks that fall under Computer Vision. And there are a variety of tools that we use for different aspects. There are tools for labeling, for simple processing, or manipulation. Typically, we use deep-learning models in these tasks to process image data. Deep-learning frameworks, like PyTorch and TensorFlow, allow us to build state-of-the-art Computer Vision models. This is technical knowhow that businesses may not have in-house. When deciding the best way to handle this, AutoML is a viable option that should be explored. There are variety of ready-made solutions that provide a layer of obstruction from these deep-learning models. We are going to explore one of these solutions in this chapter, Azure Custom Vision. Now, Azure Custom Vision exists at the Microsoft Azure platform, and we will need an Azure account and subscription in order to use Custom Vision. To do that, navigate to portal.azure.com, and create a new account, or sign into your existing account, if you already have one. Now, I already signed into my account. Once you sign in, you should be directed to your own page where you can sign up for a free trial offering $200 worth of credit. This will be more than enough to cover the content of this course. I've already set up my account, so you can see my own subscription here. I use a different subscription, a pay-as-you-go subscription. So I will move on to creating the resources that we need. We will need to create a resource group. This resource group will contain all the resources that we will use. So I've clicked Resource Groups, and then I will find the Create option in this window. Create. Okay, in this window that opens, ensure that you select the right subscription, and then we'll choose a name for our resource group. Let us choose afre fashion after the dataset that I will introduce later on. You choose a region for your resource group. You can choose USA, this is just fine. Ideally, you want to choose a region as is closest as to where your dataset sits, and then you can decide to include any tags that can identify your resource group. And then once you are done, evaluation is done and then you can click Create. Once you do this, as you see notification telling you that your resource group is being set up. I have already set up my own resource group. So you can see my afrefashion resource group, my pay is one Pay-As-You-Go subscription in the US East US location. Once your resource group is done created we are now ready to start using Azure Computer Vision.