Building a Machine Vision Mobile App to Classify Flower Species

In this chapter, we are going to use the theoretical knowledge we have learned in previous chapters to create a mobile application that will classify a specific species of flower. By utilizing use your mobile camera and pointing it at a flower, the application will analyze the image and make its best educated guess as to the species of that flower. This is where we put to work the understanding we have gained about the workings of a convolutional neural network (CNN). We will also learn a bit more about using TensorFlow as well as some tools such as TensorBoard. But before we dive in too deep, let’s talk about a few things first.

Throughout this chapter we use terms that may not be familiar to all, so let’s make sure we’re all on the same page as to what they mean.

In this chapter, we will cover...

CoreML versus TensorFlow Lite

In the machine learning world, there are two efforts (as of the time of this writing) taking place in order to improve the mobile AI experience. Instead of offloading AI or ML processing to the cloud and a data center, the faster option would be to process data on the device itself. In order to do this, the model must already be pre-trained, which means that there is a chance that it is not trained for exactly what you are going to use it for.

In this space, Apple’s effort (iOS) is called Core ML, and Google’s (Android) is called TensorFlow Lite. Let’s talk briefly about both.

CoreML

The CoreML framework from Apple provides a large selection of neural network types. This allows...

Datasets for image classification

For our flower classification example, we will be using the University of Oxford's Visual Geometry Group (VGG) image dataset collection. The collection can be accessed at http://www.robots.ox.ac.uk/~vgg/data/.

The VGG is the same department that won previous ImageNet competitions. The pretrained models, such as VGG14 and VGG16, were built by this department and they won in 2014 and 2016, respectively. These datasets are used by the VGG to train and evaluate the models that they build.

The flower dataset can be found in the Fine-Grain Recognition Datasets section of the page, along with textures and pet datasets. Click on Flower Category Datasets, or use the following link to access the flower datasets from VGG, http://www.robots.ox.ac.uk/~vgg/data/flowers/.

Here, you can find two datasets, one with 17 different species of flowers, and the...

Summary

In this chapter we have accomplished a lot in this small chapter. We began the chapter with understanding the various datasets that are available for image classification, as well as how we could obtain or create images if we could not find any that met our requirements. Next, then divided the chapter into two distinct sections. In the first section we learned about creating our own image dataset. In the second section we learned how to use TensorFlow to build the model.

In the next chapter, we are going to extend our TensorFlow knowledge even further by using various TensorFlow libraries to build a machine learning model which will predict body damage done to a car.