Evaluation of Several Artificial Intelligence and Machine Learning Algorithms for Image Classification on Small Datasets

Abstract

In this paper, a benchmark of machine learning (ML) algorithms for single-label image classification is proposed and evaluated on a small dataset. The dataset is obtained through a mobile application allowing citizens to upload images related to water and electricity distribution infrastructure problems. The collected dataset is preprocessed, organized and used to train and evaluate classical supervised ML algorithms (SVM, NB, DT, KNN and MLP) along with deep learning methods (CNN and transfer learning). Data augmentation and fine-tuning techniques are explored to handle the overfitting problem. Conducted experiment results show the effectiveness of transfer learning with data augmentation and fine-tuning using the VGG16 network as the precision reaches 89%.