Comparative Study of Machine Learning Supervised Techniques for Image Classification Using an Institutional Identification Documents Dataset

Abstract

The growing use of machine learning algorithms for image classification process using open source libraries promotes the identification of data sets to be applied on different contexts. In this way, the most outstanding algorithms for image classification with machine learning supervised techniques were implemented and their accuracy level was compared in an uncommon context, such as identification documents (combining text and images) of a university institution as a study case. The used dataset has a high complexity level in terms of design, diversity and population density. The study has provided outstanding results such as: 1) the implementation of the Transfer Learning / Image Retraining technique for which the expected performance was obtained in shorter times compared to the other techniques or algorithms studied, and 2) the need to implement a previous classifier whose objective is to refine the dataset to promote higher precision levels for new instances (new documents to classify).