Random Optimization and Entropy-Based Data Augmentation for Image Classification and Analysis “ROEDA”
Boudouh Nouara, Bilal Mokhtari · 2024
To address the challenges posed by limited and unbalanced data, data augmentation serves as a primary solution before performing machine learning techniques, mainly lassification. It allows for enhancing model accuracy, which has been proven repeatedly. This study introduces a novel image augmentation method based on a random optimization technique, referred to as ROEDA. It enhances image datasets by selecting the most dissimilar images from a set of generated variations of the original image. These variations are created by applying various filters that remove specific rows and columns from the original image.The selection of the most relevant images is achieved through an entropy-driven random optimization function, which measures the content dissimilarity between the generated images and the original. By incorporating these carefully chosen images into the dataset, ROEDA aims to diversify the training data and improve model accuracy. Our proposed method was rigorously tested using the Kaggle Cats vs. Dogs dataset. The results were promising and compelling, with our approach surpassing both the original dataset and randomly augmented datasets. The VGG16 model trained using our technique achieved an outstanding training accuracy of $\mathbf{9 2. 2 3} 1 \%$, clearly demonstrating the superior performance of our image augmentation approach.