Problem

For my academic project in the deep learning course, I was tasked to create an image classifier to accurately classify images of actresses into three categories : images corresponding to actress Keira Knightley, images corresponding to actress Natalie Portman and others. The challenge is that the faces of these two actresses are very similar so I had to work on state of the art techniques of computer vision models to accurately build this classifier.

Solution

I decided to use transfer learning and finetune a ResNet50 model pretrained on the VGGface dataset. This approach allowed me to leverage the pre-existing knowledge of the model while still being able to adapt it to my specific use case.

To further improve accuracy, I used SOTA zero-shot detection Grounding DINO model as feature extractor. This helped me achieve an accuracy rate of 94% on the test image dataset.

Impact

Achieved accuracy rate of 94%
Reduced time and resources required for image classification

Improved Actress Image Classification with CNN