Consulting

The project consisted in the development of several signal processing and machine learning based algorithms to clean, segment and classify ECG signals. The most original part was coming up with a fast method that allowed the physician to easily identify important episodes, using limited hardware. The device is produced by Eccosur S.A. (Argentina).

This project addresses the challenge of improving segmentation and classification of fruit defects (strawberries and almonds) from batch images. Accurate characterization of produce is essential to improve the distribution of premium batches, to avoid problems due to fruit diseases and to reduce food waste. From a technical point of view, the solution solved some issues related with the low amount of tagging, interpretability and variance across devices. The device is produced by AgShift (USA).

To monitor blood pressure for a long time, a cuff that inflates and deflates in a controlled way is placed on the patient arm for more than 24 hours. In this project, methods were developed to control the cuff, generating better quality pressure signals and requiring less measurement time in each inflation. A pressure signal processing stage was proposed to improve pulse detection. Also, a method based on genetic algorithms was used to optimize the model parameters. A final stage was incorporated to estimate the confidence on lost pulses based on the patient's history. The device is produced by Eccosur.

A preliminary image stegnography method was developed based on deep learning and frequency analysis techniques. The goal is to incorporate information that is not visible to the naked eye in printed commercial logos (what is known as smart-packaging, Cartocor). This way, users can retrieve information from the product logo using an app. In order to have commercial impact, the coded information must be barely visible. A major challenge is to achieve watermarks that are legible once printed, something that classical stenographic techniques fail to do.