Some of my projects on computer vision.
All these are non-guided projects.
Project built using a compilation of different data sources all around the world with more than 16,000 lungs radiography.
Techniques used:
- Modeling
- Batch training
- Fine tuning
- Transfer learning
- Class activation map
The model created is able to correctly diagnose with an accuracy of 95.25% on the test set (99.98% and 98.4% on the traning and validation sets respectively. It's also able to localize where the CNN was able to see the patterns to classify each radiography as covid negative or covid positive.
Not ComputerVision per se, but a very important part of it as in this project analyzes a 3D image of a rock, segmentates it and analyzes its porosity. Techniques used:
- Image analysis
- Segmentation
- Porosity calculation
3D representation of the pores:
Building a classificator that detects 20 different characters from The Simpsons.
Techniques used:
- Convolutional Neural Networks modeling
- Batch training
- Data augmentation
- Fine-tuning
- Transfer learning
It's a quite challenging project given the nature of the images, every character is presented in a variety of situations with different backgrounds, different angles, different facial expressions and sometimes even different outfits.
It was also interesting to see how different models reacted to external data, such as photo-realistic depiction of The Simpsons characters: