Training For Facial Recognition

Training For Facial Recognition

To learn more about what Facial Recognition Technology is, please read our article.

As with all scOS AI technology, it begins with two things; data and training.

Data, in the form of images, are required to show the potential AI model examples of what it needs to find by itself.

Therefore, the purpose of this article is to show how scOS trains AI on how to see a person and face within a single image - like you would when going about your day.

Due to the vast amount variations in an individuals face compared to the next, many thousands of images are required when compared to ANPR Training. As an example, the license plates on cars are very consistent with their features. However, this is not the case for individuals.

Image Collection

As we are working on the initial stages of scOS, we have to be creative on how we responsibly collect imagery. Furthermore, we also have to consider a demographic balance of the people in the collected images to prevent any technical bias.

We have two methods for image collection:

Method One: Consenting Customers

When we need to gather imagery, we ask our customers if they are happy to provide us with training images that contain people on their property. They choose what images they send to us and how many.

This is the best data to receive as it is most accurate for our use case on CCTV cameras.

However, because we are still in the testing stages of our organisation, we only have a handful of customers who can provide imagery. Therefore, we have to source more.

Method Two: Free Commercial Use Images

We use freely avilable stock photos that are suitable for commercial use. The main and only website we have used is Pexels, where publishers understand their content is published on the internet and is permitted to be used for free. Here we can use search terms like "people", where Pexels will return every* image on the website that contains people.

We then download these images and tidy up to remove duplicates or images not fit for purpose (ones that don't include people).

* This will be every image that has been tagged with the word "people".

Image Processing

These images are then placed in batch folders and securely sent to team members for annotation.

Image annotation is where a team member looks at a single image and draws a box around the "object", in this case it is a person and face. For example, imagine you have a portrait of a person, you want to draw a box around the person so they are inside it. This box will be labelled "person". A second box is then drawn around the persons face. This is labelled "face".

Later when the AI model is shown an image it has never seen before, it will repeat the same steps you just did on its own accord. Thus it now knows what a person and a face looks like.

As humans are biological, we have a very natural learning process to this. However, with AI, we have to be precise on what we are getting it to recognise.

scOS Sample Annotation

The image above shows exactly what the annotation process looks like. This image is freely and publicly available on Pexels.

Additional Notes

ONE

Within the "person" label, scOS also includes held objects within the annotation. This is so, later, our AI model can associate an object being held in a persons hand.

There are two clear use cases for this:

  1. Is an unknown person holding a dangerous item, will this make them suspicious?
  2. This messenger bag is being carried by this person, could it become lost property?

TWO

To assist with annotation speed and to help with ethics and privacy, we may use the already trained AI model to annotate new images.

Model Training

Once all the images have been annotated, we will have two new files; the image and its corresponding data file that holds the coordinate locations of the boxes.

As these images are used for training our object detection model "scOS ObjecTron", they are mixed into a new dataset with the ANPR and vehicle annotations.

scOS ObjecTron is the technology that interprets what is visible on your camera.

Once training is complete, we then asses the results of the model by testing it on images it has never seen before.

DATA RETENTION:

Given the extensive efforts required to gather a comprehensive dataset encompassing the necessary variations, it is essential for us to retain the collected images while continually improving our model. The retention period is limited to no longer than two years from the date of collection, thereby striking a balance between model enhancement and adhering to responsible data management practices.

INDIVIDUAL RIGHTS:

We respect the rights of individuals under the GDPR. Users possess the right to access their personal data held by us, request rectification of any inaccuracies, and request the deletion of their data. For exercising these rights or any inquiries concerning data processing, individuals may contact us at controller@smartcityos.co.uk or reach out to us by phone at 020 8050 1095. Prompt responses will be provided, and appropriate actions will be taken in accordance with applicable data protection laws.

CONTACT INFORMATION:

For any queries or concerns regarding data protection and privacy practices, individuals can contact us at controller@smartcityos.co.uk or call us at 020 8050 1095.

Please also view more readable policies and articles around data in our Data Ethics Zone.