Projects
Spectrum photo editor
Android application that converts regular photos to artistic images using Neural Style Transfer(K-style version). The neural network runs completely on-device, leveraging the power of depthwise-separable convolutions. I used tensorflow-lite as the inference library and have also added GPU-support for supported mobile-devices.
Available on Google play
Signature verification
Writer independent off-line Signature verification using Siamese Convolutional networks. Trained the model on the CEDAR1 dataset and achieved a AUC of 75.34%.
View Project
Cancer detection and Diagnosis - Mylan hackathon '18
Made for Mylan Hackathon 2018 organised by Mylan Pharmeceuticals. The project is a flask website for showcasing some of the machine learning models we built for early detection of cancer through lifestyle habits and non-invasive tests. We used Convolutional Neural networks for diagnosing cancer through patient's tissue samples. We also built a model for binary and multiclass classification of chemotheraphy drugs(protein kinases) using their drug fingerprints.
View Project
Image color quantization using K-means
Implemented the K-Means algorithm in C++ to convert a given RGB image into a K-bit RGB image, thus quantizing all colours in the image to the K most dominant colours(hence compressing the image file size in the process).
View Project
Denoising Autoencoder
A denoising autoencoder trained on the RENOIR dataset. Wrote the model in Tensorflow(layers API).
View Project