Synthetic data generation and vision AI model training for book recognition

Pocketbook "Introduction to Machine Learning with TensorFlow.js" is the absolute beginners' guide to Machine Learning for JavaScript developers. This book will teach you the essentials for Machine Learning using the JavaScript version of the world's most popular Machine Learning framework, TensorFlow.js. The book was published in February 2021 by Asim Hussain.

Our team worked with Asim Hussain to generate a computer vision AI model using Synthetic data.

Synthetic data generation

With the use of realistic 3D models and the help of syntheticAIdata service you can train your custom vision model with 3 simple steps:

1. Uploading the 3D model
2. Configuring the options
3. Generating Synthetic Data

Our team prepared a realistic 3D model of the book and texture for the model. Then, they chose background images from our “Indoor” image category. During the model generation process, the 3D model of the book was rendered from multiple angles. This strategy allows vision AI to detect the book from different angles, which increases the model's accuracy.

Once we generated the images, we used the integration feature to Azure, which automatically uploads the images to Azure Custom Vision. In Azure Custom Vision, these uploaded images are tagged correctly and ready to use for training train the model.

The following video shows several synthetic images generated in the process and testing results.

Object detection

Detection of books with vision AI could be a good fit for multiple use cases, such as:

- Identifying the most-sold book in a bookstore
- Detecting which books are being borrowed or returned in a library
- Verifying the printing process in a printing office
- And many more

Creating a 3D model of the book is simple and fast, and the graphics that were previously used for printing the book cover can be re-used for the texture. Generating synthetic images for vision AI model training shortens the time and minimises costs for data collection.

Contact us via hello@syntheticaidata.com if you have any questions about this experiment or whether you want us to help your business with our product. We are happy to help!

About Asim Hussain

Asim Hussain is the director of Green Software and Ecosystems at Intel and Chairperson at Green Software Foundation. He is passionate about sharing his knowledge with others. He published many interactive JavaScript courses for free on Code Craft.

He often talks about Machine Learning, JavaScript, Sustainable Development, and Developer relations at conferences. Some of the recordings are available on his website.