Herta looks to Deep Learning techniques to improve facial recognition
Axis partner Herta develops and designs facial recognition software solutions. It is specialized in the simultaneous detection and identification of multiple subjects in crowded scenarios such as airports, train and metro stations, sports stadiums, shopping malls. It offers solutions for videosurveillance, access control and marketing requirements.
We had a chance to sit down with Herta Marketing Executive Laura Blanc Pedregal, to talk about how they are using deep learning techniques to improve facial recognition.
How does deep learning apply to your core business and products?
Deep learning techniques are currently state-of-the-art in fields like computer vision and speech analysis. Herta’s early adoption of deep learning has dramatically improved the accuracy of most of our algorithms, making them robust to the most challenging and complex scenarios.
What is the difference between deep learning and Artificial Intelligence?
Artificial intelligence, as coined 60 years ago, seeks to provide machines the ability to perform rational tasks, to exhibit “intelligence”. In a general sense, this meant creating machines with comprehensive knowledge and cognition, indistinguishable from a human. In the narrow sense, it means developing technologies that can perform specific tasks as well as humans, or even better, by making autonomous decisions.
Machine learning aims at creating algorithms that learn from observed data, so that they are able to make determinations or predictions over new, previously unseen data. It is often considered a subset of AI, its learning component. Deep learning is a branch of machine learning. It is particularly useful for certain learning tasks, and it benefits from large amounts of data.
How far away are we from seeing deep learning being utilized in commercial applications?
Deep learning is already in commercial use. We have recently launched the new version of all of our products, which are now all based in deep learning. It is very close to become a real product, because it is already working. Deep learning has helped us improving the performance of our software to levels that were unreachable a few years ago.
How long does it take for deep learning to become really effective in any particular installation – does it get better over time?
This is a common misunderstanding on how the technology is deployed. The training process is challenging and computationally intensive, so it is carried out behind the scenes.
It takes days or even weeks of intensive computations over large amounts of carefully annotated data, and produces a trained model as a result. During deployment, the trained model makes predictions from new observations, but in general it is not updated based on these — the model remains unchanged.
Currently, updating the model would require the new data to be also precisely annotated. This is important to avoid potential degeneration of the model, which would not be acceptable in real environments.
Does technology like this raise privacy concerns?
The fact of choosing deep learning over alternative learning paradigms does not make any difference in terms of privacy. The internal representation of a face cannot be linked back to the image that originated it. Privacy concerns must be focused on database management and video capture, when faces are still images, so before deep learning takes over.
Other than security, how else can advanced analytics such as this be beneficial to a company or organization? What other use models have you discussed with customers? (e.g. customer targeting, traffic analysis/floor planning, etc.)
Deep learning is particularly important for our solution of gender, age and ethnicity recognition, because it allows a considerable accuracy increase due to the possibility of holding millions of images in order to build a model in the most effective way.
We have also noted many other uses in terms of analytics such as people counting and tracking. It is very important to remember that customers are key for every business and having the tool that allows you to know their behavior and characteristics is fundamental to find the best way to achieve them.
How difficult would it be for an average end user – say a retail store – to install, use and derive meaningful benefits from this type of technology?
Our technology is very easy to install and implement, like plug-and-play. You just need to install the software and start using it. We have made sure to design a software that is user-friendly and intuitive so that any end user can make use of it.
When it comes to benefits we can talk about a wide range of advantages you can find when implementing this type of technology, from higher security to obtaining meaningful data such as your client profile, number of clients or the hot-spots of your business.
Laura Blanc Pedregal, Marketing Executive, Herta Security
Laura is Licenciate in Business Management and Administration by the University of Barcelona (2013) and during her studies she joined the Erasmus Project in Utrecht (The Netherlands) where she learnt about International Business & Management (2012). She holds two Postgraduate Diplomas in Marketing & Market Research (2014), and Neuromarketing (2015), and she is currently studying a Master in Marketing Management at EADA business school (2017). Prior to Herta, she worked at blur Group (London and Exeter, UK) in the Digital Marketing team (2013). In 2014 she joined Herta Security in the marketing department, where she works in promoting the company and its products, as well as supporting the sales team.