Region of Interest Detection Using AI Techniques and Morphological Operations

Abstract

Nowadays, with the increase of technological advancement, AI can be found anywhere, from smartphones with facial recognition, voice assistant, and text autocorrect, to security systems to detect unauthorised persons, cyberbullying, cyberattacks, or even in healthcare to assist doctors in giving diagnostics and treatments. The intersection point of all these use cases is that they all need a lot of labelled data to obtain good results in classification processes. Labelling is tedious work that can consume many resources, like time, workforce, computer power, and money. The labelling process differs according to the type of data that you are labelling (audio, video, images, text). This paper focuses on the image type of data. To solve this issue, the paper proposes the use of an auto-labelling algorithm by employing a combination of a Convolutional Neural Network (CNN)architecture and morphological operations with a set of hyperparameters. To quantify how well the algorithm performs, an efficiency metric given by two components, a local evaluation of the results and a global evaluation of the results, is established by comparing the labels given by the algorithm with the labels manually done. Furthermore, the paper analyses the impact of the hyperparameters on the algorithm using the established efficiency metric. Finally, the results for the best configuration of the hyperparameter are explored using both global and local data analysis. Using the efficiency metric given by the overlap area, the algorithm achieved a mean overlap percentage of 89.33%.