Hi there 👋
My name is Luca and I'm currently working on a custom CRM to enable recommendation based on a Deep Contextual Multi-Armed Bandits algorithm. In the past I used to work a lot with images, automated processes and OCRs. In order to solve some of the problems I tackled, I had to create or fork a couple repos that you can find here. In particular:
Shutter
My first side project, that then became my python playground to test some programming approaches.
It started as a tool to benchmark image preprocessing algorithms against OCR implementations.
Needs changed over time and eventually became a tool to generate document images to train deep learning models.
I used it to pretrain models with different architectures and then finetune on the actual problem I had with a limited dataset.
I successfully trained a classifier for documents and a Super Resolution GAN to improve image scans, I was planning to train a detector and a custom OCR as well but priorities have changed.
EfficientNet-PyTorch
Forked the original repo a while ago and modified it to be trained on single channel grayscale images; nothing special but worked like a charm (94,5% acc on rvl-cdip) / (98% acc on my classification problem)
Doc-SR
More or less same story here. I had a lot of images to pass to an OCR algorithm that failed to produce a consistent result due to poor quality of the scans... Eventually ended up with this fork which I used to train a super resolution GAN to mainly do 3 things:
- Denoise Image
- Invert White on black text
- Generate a super resolved version of the original
Results were pretty neat and granted me up to 500% accuracy gain on OCR tasks that were hard failing without this preprocessing. It was all about generating the right images with shutter and add the right loss functions.
Check this result
Botox
Maybe the coolest one, a Mac OS tray app thought to easily insert your username and password every time you need to (I used to do it 8/10 times a day!). It can be invoked from the touchbar (Quick Action) or with a keyboard shortcut.
The credentials are kept encrypted and decrypted only inside the stack of the functions doing the typing job.
100% secure? maybe not. Handful? definetely. Say goodbye to keychain access popup or login screens.
Booko 🤖
A telegram bot to help you filter tennis fields availability around you (Sometimes it takes a bit to wake up from naps imposed by the hosting service 💤) Try it
Contact
'@'.join(['lucaranalli7', 'gmail.com'])