nik1806 / sober_drunk_classification

The project aims to build a classifier that can classify for an infrared image as sober or mild drunk or drunk.

Task

The task is to build a classifier that can classify for a given image as sober or mild drunk or drunk. In order to train or test, u can consider all four, front, side etc. at once to classify whether drunk or sober.

About dataset

Attached is the data of 41 people, taken from IR sensor. For each person, there are 4 types of images taken: sober, 20 mins after drinking, 40 mins after drinking, 1hr after drinking 4 glasses of wine. For each type for each person, 4 images are taken: front face, side face, eyes, and hand palm.

Installing dependencies

I have used virtual environments for handling the dependencies. Run the following command:

pip install -r requirements.txt

About the repo

• This repository explores and presents 2 different approaches to tackle the problem.

1. Using transfer learning - based on concepts of deep learning - achieving an accuracy of 87.5% (More details inside drunk_classification_dl directory).
2. Using rudimentary methods of feature extraction and separate classifier - based on classical machine learning - achieving an accuracy of 58% (More details inside drunk_classification_ml directory)

sober_drunk_classification's Issues

Need to utilise all four different images

Incorporating all 4 sub-class images to classify to further improve the accuracy.

One way could be merging images. (it won't be able to increase the dataset length)

another could be to use multiple classifiers and join the results at the end using FC layers. (Something similar to Adaboosting)