Comparative Analysis of Classification Models Using Midjourney-generated Images in the Realm of Machine Learning

Abstract

Artificial intelligence (AI) integration has shaped rapid and remarkable advances in machine learning. In the relentless pursuit of advancing AI capabilities, applications such as Midjourney emerge as pioneering tools designed to create intricate images from the essence of textual prompts. Midjourney, an exampleof a generative AI tool, utilizes text-to-image methods with an extensive database. This study aims to provide insights into the potential advantages and limitations of generative AI images in machine learning. This research methodology explores Midjourney to generate 500 images of dogs and cats. Subsequently, these images serve as the basis for building classification models. We will explore the classification models and their evaluations in three scenarios: i) 100% of images generated by Midjourney, ii) 100% of real images, andiii) 50% of images generated by Midjourney and 50% of real images. To achieve this goal, the study utilizes twocommonly used deep learning models, InceptionV3 and EfficientNetB4, for training and testing the classification models. The analysis results indicate a significant improvement when combining generated Midjourney images and real images for classification. This comparative examination high lights the effectiveness o fAI-generated images in enhancing the performance of machine learning models, emphasizing the potential to augment the image subset with synthesized images from generative IA.