Before we dive in too deep, let us first talk about a term you will hear used quite a bit in this chapter, MobileNets. What is a MobileNet you might ask? Simply put, it’s an architecture which is designed specifically for mobile and embedded vision-based applications. On such devices there is a lack of computing power available for such processing, which therefore increases the need for a better solution that one used on a desktop environment.

The MobileNet architecture was proposed by Google, and briefly:

1. Uses depth-wise separable convolutions. This significantly reduces the number of parameters when compared to a neural network using normal convolutions with the same depth. The result is what is known as a light-weight deep neural network.
2. Depth-wise convolution, followed by Pointwise convolution, replaces the normal convolution process.

In order to simplify things, we are going to break down this chapter into the following two sections:

• Datasets for Image Classification: In this section we will explore the various datasets (all of which are available online) that can be used for image classification. We will also address the issue of how to create our own datasets, if necessary.

• Using TensorFlow to Build the Model: In this section we will use TensorFlow to train our classification model. We do this by using a pretrained model called MobileNet. MobileNets are a family of mobile-first computer vision models for TensorFlow, designed to maximize accuracy while considering the restricted resources available for an on-device or embedded application.
• In addition, we will look at converting the output model into a .tflite format, which can be used within other mobile or embedded devices. TFLite stands for TensorFlow Lite. You can learn more about TensorFlow Lite via any internet search engine.