Classificação de imagens por relevância das partes usando machine learning

Abstract

The field of Artificial Intelligence is rapidly growing, with Deep Learning techniques standing out. These solutions aim to improve processes, despite high computational costs. This work proposes the use of classic Machine Learning algorithms, such as KNN and SVM, to optimize image classification. The study adapts a relevance technique, originally used in audio classification, to an animal image classification problem. This technique assesses which parts of the image are most important for the model’s decision, aiming to eliminate backgrounds and objects outside the class of interest without the need for prior segmentation of these elements. Finally, the results are compared with a traditional approach and with the probability summation technique. When applied to an animal classification database, the result with the technique achieved an F1-Score of 94,4% with SVM. In addition, the relevance of the regions is highlighted through a heatmap, which gives the technique the ability to explain the decisions made by the classifiers. With the increasing complexity of models, understanding how they make their decisions through explanations helps confer greater credibility, transparency, and trust in these technologies.