Design of a Generative AI Image Similarity Test Application and Handmade Images Using Deep Learning Methods

Abstract

Purpose: The aim of this research is to develop a classification model using the Transformer approach, specifically the BEiT architecture, to differentiate between handmade images and AI Generative Art. The objective is to ensure the authenticity of art and address ethical and legal concerns related to AI Generative Art.

Design/methodology/approach: The study utilizes the BEiT architecture within the Transformer approach to create a classification model. The training process uses Bidirectional Encoder representation from Image Transformers (BEiT) to improve image classification. The primary datasets are collected through a Python image scraper program. The BEiT workflow includes Pre-training, Masking, Inpainting, and Interface Design with Gradio.

Findings/result: The Transformer model, using the BEiT architecture, achieves 96.34% accuracy and 0.0921 loss in differentiating handmade images and AI Generative Art. The model demonstrates a balanced precision and recall in each category, outperforming previous methods such as Convolutional Neural Network (CNN) and VGG16. The language used is clear, objective, and value-neutral, with a formal register and precise word choice. No changes in content were made. The Gradio interface was used to successfully test the model.

Originality/value/state of the art: The research presents a state-of-the-art classification model that uses the Transformer approach, specifically the BEiT architecture, to differentiate between handmade and AI Generative Art images. The research presents a state-of-the-art classification model that uses the Transformer approach, specifically the BEiT architecture, to differentiate between handmade and AI Generative Art images. The text adheres to conventional structure and formatting features, including consistent citation and footnote style. The sentences and paragraphs create a logical flow of information with causal connections between statements. The text is free from grammatical errors, spelling mistakes, and punctuation errors. Additionally, the research is enhanced by the innovative approach to data collection using a Python image scraper program.