Projects
Find For Me: AI-based app to find alternatives for your desired luxury apparel
- Developing a client-server web app for visual search/match of luxury clothing alternatives from fast-fashion retailers using Python, Flask, and JavaScript.
- Collected a dataset of 190k fashion items from online stores through distributed web-scraping.
- Deployed app to production on Google Cloud Service, optimized ML system, and server configuration, resulting in a memory reduction by 40% and cost reduction by 30%.
Deep Learning for Dental Hyperspectral Image Analysis
- Built a deep learning-based pipeline in Python, Keras, and TensorFlow to segment disease areas from hyperspectral images of oral cavities reaching IoU segmentation score up to 0.92.
- Developed a real-time data generator for hyperspectral image augmentation; image segmentation and visualization tools for hyperspectral images based on Unet and cloud computing.
- Published a first-author scientific paper which was awarded 2nd place for ”Best student paper award” (among approx. 200 participants).
Automatic quality control of fruits and vegetables by surface color appearance in the growing stage
- Built a system for the automatic detection of a color checker in a natural environment under varying illumination with an accuracy of 96%.
- Applied color correction to exclude the effect of the camera and illuminant; evaluated the color correction accuracy.
- The project was awarded a prize as the best-proposed solution, suitable for use in the industry.
- Software: Matlab, Image Processing Toolbox.
Counterfeit banknotes detection using Hyperspectral technologies in UV, visible and IR ranges
- Captured hyperspectral images of real banknotes under different illumination (UV, VIS and IR ranges).
- Applied PCA, ICA to identify spectral bands with unique features.
- Developed a proof-of-concept solution for fake banknote detection based on detected features.
- Software: Matlab, Image Processing Toolbox.
3D reconstruction of scene from 2 Kinect cameras
- Captured 3D scene by using a pair of Kinect V2 cameras.
- Combined RGB-D images to reconstruct the corresponding 3D scene.
- Software: Matlab, Image Processing Toolbox.
Just for fun
During the project, our team developed a remotely controlled robotic system with cameras and lasers to map the Mars area and create the shortest route between the objects. Our team became a winner of the final competition, providing the most accurate mapping among other teams and reported the shortest route.
As a team leader, I was responsible for task management and team building between people from a different scientific and cultural backgrounds.
My Pill
During the project, our team developed a new business model related to Color Science Branch. The idea was to create a device which contains NIR spectrometer for evaluation of the true composition of medicine before consumption to avoid low-quality, fake or substandard pills.