Device enables computer to identify whether user is male or female

The Spanish Patent and Trademark Office has awarded the Universidad Politécnica de Madrid and the Universidad Rey Juan Carlos Spanish patent ES 2 339 100 B2 for the device.

Thanks to the new algorithm, devices can be built to measure television or advertising video audiences by gathering demographic information about spectators (dynamic marketing). The new device is also useful for conducting market research at shopping centres, stores, banks or any other business using cameras to count people and extract demographic information. Another application is interactive kiosks with a virtual vendor, as the device automatically extracts information about the user, such as the person's gender, to improve interaction.

A step forward in gender recognition from facial images

This research, the results of which were published in IEEE Transactions on Pattern Analysis and Machine Intelligence, demonstrates that linear techniques are just as good as support vector machines (SVM) for the gender recognition problem. The developed technique is applicable in devices that have low computational resources, like telephones or intelligent cameras.

The study concludes that linear methods are useful for training with databases that contain a small number of images, as well as for outputting gender classifiers that are as fast as boosting-based classifiers. However, boosting- or SVM-based methods will require more training images to get good results. Finally, SVM-based classifiers are the slowest option. Additionally, the experimental evidence suggests that there is a dependency among different demographic variables like gender, age or ethnicity.

Device for demographic face classification

The invention is a device equipped with a camera that captures digital images and is connected to an image processing system. The image processing system trims each face detection image to the size of 25x25 pixels. An elliptical mask (designed to eliminate background interference) is then applied to the image, and it is equalized and classified.

The device advances the state of the art by using a classifier based on the most efficient linear classification methods: principal component analysis (PCA), followed by Fisher's linear discriminant analysis (LDA) using a Bayesian classifier in the small dimensional space output by the LDA. For PCA+LDA to be competitive, the crucial step is to select the most discriminant PCA features before performing LDA.

One of the major research areas in informatics is the development of machines that interact with users in the same way as human beings communicate with each other. This research is a step further in this direction.