As part of Centre for Intelligent Systems, Department of Computer Science, TalTech, I am continuously involved in projects involving data science and machine intelligence. Here are a few projects that I’ve been developing software for.
Development of artificial intelligence methods based system for classification and risk groups determination of motor insurance clients (2016)
An R&D project done for a large insurance company. My role in it was twofold:
- Data (pre)processing. Mostly textual data and mostly in MATLAB. Further analysis and artificial neural network training were done in Python environment.
- Putting together the final deliverable—a graphical application for Python which included data processing and application of pretrained neural networks to obtain the predictions according to the project goals.
For me this was the first project where I would employ a MATLAB and Python co-development cycle.
A Methodology for computerised detection of pavement cracks and other road defects (2018)
In this project
, my task was conducting a research of deep learning methods for image recognition and the implementation of corresponding algorithms to the problem of detection of pavement distress from nonideal photographic images of highway roads. This was a so called smart specialization project and research and development was done for a company that deals with issues related to highway road inspection.
This was a challenging project not only because I had little previous experience with machine vision but also because of the essence of the problem. The photos of highway roads were of varying quality feature-wise oftentimes corrupted by shadows and random roadside objects making it to the final composite photograph.
Thus, several issues had to be tackled. In terms of computational intelligence algorithms, convolutional neural networks were finally considered as the tool for image classification—the output of the complete network predicted the presence of a defect on a square-shaped segment cut from the large photo based on a grid (pixel resolution of the photos was considerably high—another challenge to overcome).
Other team members also tried various machine learning methods, but convolutional networks produced the most coherent results in the scope of the project. In the end, precision/recall metrics were still low, but the overall system was already capable of detecting defect-like objects on arbitrary images.
The deliverable—a frontend completely written in Python and PyQt5—was then handed over to the company. Some backend functionality (especially parts related to initial image mask manipulation and road extraction) was co-developed with Dr. Andri Riid.
Also, the project has grown into a bigger one and is currently ongoing.