Overview / Usage

The aim of this project is to create a robust system that can take a front faced image of a person as to its input and can predict his(s) age. The challenges lie in the data itself. One of the key problems to address here is how can we make a computer understand that BW images do not necessarily mean that the age of a person present in them is old. The dataset used for this project suffers from the terrible problem of label noise which means some of the labels (which in this case is AGE) of the images are wrong. So, a sufficient amount of data cleaning is required in order to develop this project.

Here is a demo of the web application which takes a front faced image of a person and predicts his(s) age: https://www.loom.com/share/587a722d8a644ceeb80748a28356109f

The main deep learning model is also exposed as a REST API via FloydHub (can be made public if necessary). Here is a demo of how to consume the API in Postman: https://www.loom.com/share/0665aad786b7488bb3bbf1c4cb2e38c5

Methodology / Approach

• Used a pre-trained deep architecture named RestNet50
• Used one cycle policy to enhance the effectiveness of model training
• Used manual annotations to clean the data
• Used image similarity measurements to remove duplicates
• Used data augmentation to help the model generalize well
• Used mixed precision policy to reduce the training time

Technologies Used

Python (Language) (Intel's version)
fastai (Deep Learning library)
Google Cloud Platform (model training)

Repository

https://github.com/sayakpaul/Age-Detection-of-Indian-Actors