USING MICROSOFT CUSTOM VISION AI TO BUILD A TOOL

Last year marked another milestone in my professional journey when I completed the Microsoft Azure AI Engineer Associate certification. Excited to apply what I had learned, I decided to dive into the world of Azure Custom Vision AI tools and APIs. My goal was to create something practical and impactful, and that’s when I came up with the idea of developing a tool integrated with our Car Park Digital Signage Module. This tool would be able to identify cars and their number plates—a project that turned out to be both challenging and rewarding.

The Concept

The core functionality of my tool is quite simple yet powerful. Using Azure’s Custom Vision AI and Optical Character Recognition (OCR), it can detect a car, and the car number plate to determine whether the vehicle is permitted to park in a specific location. If the car is not authorised, the tool will display a message on the digital screen, guiding the driver to park in the correct area. I felt this tool would not only streamline our client’s parking management but also enhance security for our client’s facility.

How I Built It

For the development of this tool, I opted to use Microsoft Visual Studio, complete with PHP extensions, to bring my vision to life. Additionally, I utilised GitHub to manage the code changes and repository, allowing me to seamlessly integrate with Copilot Pro for assistance during coding.

Of course, to run the tool locally, I first had to install PHP. Once set up, I began the exciting process of building out the features.

Key Features of the Tool

The tool is equipped with several features designed to enhance its functionality:

• Webcam Integration: It can use a webcam to dynamically display bounding boxes around identified objects, allowing real-time visual feedback.
• Custom Vision Model: Employing Azure Custom Vision, the tool effectively identifies both cars and number plates, making it reliable for our parking needs.
• Adjustable Settings: Users can modify brightness and contrast to improve image clarity based on different lighting conditions.
• Tag Selection: There’s an option to select specific tag names for identification, which can also be used for the OCR process.
• Stream Control: Users can start and stop the video stream, with settings to determine how frequently to capture video frames.
• Probability Threshold: A pivotal feature that allows users to set a probability threshold for object detection. A higher threshold increases the chances of correctly identifying an object. However, if the model isn’t trained with an extensive dataset, this threshold may need to be lowered for accurate detection.
• Manual Image Loading: Besides real-time video processing, users can manually upload images (PNG, JPG) to identify the tag bounding areas, providing flexibility in how the tool can be used.
• Status Updates: The tool also displays status messages, along with the date and time of any changes made, ensuring full transparency during usage.

Lessons Learned

Embarking on this project has taught me invaluable lessons about AI and its integration into our daily workflows:

1. Emerging Technology: AI is rapidly evolving and will soon become an integral part of our everyday tasks. The capabilities it offers can make our work much more efficient.
2. Data Importance: When working with Custom Vision AI, it’s crucial to use large datasets to train the model effectively. I only used about 100 images, and the project would have benefited from a larger dataset to improve accuracy.
3. Human Oversight: Despite the automation that AI brings, human oversight is essential. There needs to be a person verifying the output to ensure accuracy and reliability.
4. Abundance of Tools: There is a plethora of AI tools available that can help achieve various business objectives. It’s about finding the right tools that suit your specific needs.

One of my surprises during this process was the lack of existing models specifically designed to identify number plates. I ended up using Azure OCR for character recognition on the number plates. I also experimented with Tesseract.js; however, I found that it did not significantly enhance the recognition capabilities.

Creating this tool has been an engaging journey, blending my newfound skills in AI with real-world applications. I truly believe that this technology can shape the way we manage parking, making it smarter and more efficient. As I continue to explore the possibilities within AI, I’m excited about the potential to innovate and create solutions that make a difference.

If you’re interested in the world of AI or facing similar challenges in your organisation, I encourage you to just dive in. You may find exciting opportunities waiting at your fingertips!

If this article helped you in any way and you want to show your appreciation, I am more than happy to receive donations through PayPal. This will help me maintain and improve this website so I can help more people out there. Thank you for your help.

Marco Tran

I am an entrepreneur based in Sydney Australia. I was born in Vietnam, grew up in Italy and currently residing in Australia. I started my first business venture Advertise Me from a random idea and have never looked back since. My passion is in the digital space, affiliate marketing, fitness and I launched several digital products. You will find these on the portfolio page.

I’ve decided to change from a Vegetarian to a Vegan diet and started a website called Veggie Meals.

I started this blog so I could leave a digital footprint of my random thoughts, ideas and life in general.

If any of the articles helped you in any way, please donate. Thank you for your help.

Affiliate Compensated: there are some articles with links to products or services that I may receive a commission.