Mandela Aikaterini (Phd Candidate)

Thesis title: Development and implementation of digital image classification models for the prediction of economically exploitable mineral raw materials
Supervisor: Kotsakis Rigas
Advisory Committee Members:
Georgios Bamnios, Professor IHU
Ioannis Kapageridis, Associate Professor, University of Western Macedonia

The field of artificial intelligence has shown a rapid growth recent year, and this is largely due to machine learning and toevolution of neural networks and deep machine learning. With the term deeplearning we mean the use of multi-layer neural networks for analysis andfinding patterns in lots of data. The application of this technology varies whileevery day there are new ways to use it.

The purpose is initially to research the field of deep machine learningstarting with basic concepts and then specializing in analysis and processingimage. Research and analysis of all aspects that make up a neural network, withintended for application in computer vision, such as the classical neural network, theconvolutional network and their training process through error functions,steep descent algorithm etc.

Also, development of methods for evaluationof such models but also concepts that serve to detect objects inimages and their segmentation. Additionally, observation and evaluation of technologiesand architectural models based on the convolutional networks considered Stateof the Art in prediction by image classification. Expanding on the abovefindings in the experimental part will develop class classification models wherethey will separate the lower exploitability limit (Grade Cut-off) to produce useful raw materials yielding some profit, based on the concentration rateof the metallic elements, from the earth’s solid crust.

After the foundation is laid forunderstanding the technologies, the evolution will be to create an applicationof solving a problem of deriving quick decisions of financial importance forMineral Raw Materials (RMRs).Specifically in the analysis and processing of images from outdoors and undergroundexploitation of deposits for the purpose of first determining the limits throughsegmentation and then their classification into various classes.
Within theseclasses, various characteristics, geometric shapes,boundaries, aspects, and properties that will be their common characteristicunder study ofMineral Raw Materials(RMRs).

The aim is to implement as many as possiblepre-processing and training techniques are available to maximize performanceof the model to the input of new data. For this they should be testeddifferent parameters and versions of architectures so that they can be compared, andthe best are selected.Finally, problems can be combined to create a result of a channel processing the image as we receive it from the stageof mapping the geological front or pattern. It is worth noting that theutility of such a tool in real business applicationsoptimizes techniques commonly used today in such problemsdecisions.