Age detection


Age estimation has become relevant to an increasing amount of applications, particularly since the rise of social platforms and social media. It has various applications, ranging from customer relations to biometrics and entertainment. Since every  person ages differently, real age may not be easy to  be predicted from the face images.

Image 1: Age and gender detection on

Human face serves as a knowledge base for multiple useful information as gender, expression and age. In the case of estimating human ages, there are a wide range of variation in both shape and texture of the faces. Estimating age from images has been historically one of the most challenging problems within the field of facial analysis. Some of the reasons are the aging process, high variance of observations within the same age range, camouflage due to beardsmustacheglasses and make up. The difficulty to gather a complete and clean, sufficient training data is another problems. As most in image recognition tasks, a large and representative amount of data is required for the model to be trained successfully.


The dataset we used is from the IMDB WIKI 500+ dataset, with the image size of 224×224 as the input of our deep model. We also used 40% of the height and width of the detected face (in this dataset they use a technique that add 40% margin in the crop face, because they consider the content around the face also effect the age estimation task. So here we use the same technique).

Image 2: The IMDB WIKI 500+ dataset

The raw dataset from IMDB and WIKI is very noisy (wrong labels) so we have to omit some of the class. We have tried several different model and finally we chose to fine-tune the VGGFace deep model for our problems due to it’s performance. The VGGFace architecture had already been trained on millions of face images, we assumed that this model had already been trained to extract many of the faces features. Thus, we used a technique called transfer learning, which used the already extracted features by VGGFace and fine-tuning it on the IMDB WIKI 500+ data set, and then another fine-tuning step on a small dataset of Asian faces (we are currently focusing on Asian faces). The result is quite close to the real age (with the error of 3) and make quite steady predictions. But as we said, the problem of age estimation is very difficult through out the last few decades so we can’t expect an absolute precise results yet.

Image 3: Our model predicted results (actual age: left 24, right 29)

 Real life applications

  • Access control: Restricting the access of minors to sensible products like alcohol from vending machines.
  • Police officers: Automatic scanning of video records for suspects with an age estimation during investigations.
  • Advertisement: Smart advertisement board that adapting to its offer for young, adult, or elderly people.

Other small applications such as assessing plastic surgery, facial beauty product development, theater and movie role casting, etc.



Rasmus Rothe · Radu Timofte · Luc Van Gool, 2016, Deep expectation of real and apparent age from a single image without facial landmarks