DETECTION OF AI-GENERATED ANIME IMAGES USING DEEP LEARNING

Kusuma, Surya Widi and Natalia, Friska (2024) DETECTION OF AI-GENERATED ANIME IMAGES USING DEEP LEARNING. ICIC Express Letters, 15 (3). ISSN 2185-2766

Text DETECTION OF AI-GENERATED ANIME IMAGES USING DEEP LEARNING.pdf Download (1MB)

Abstract

Advances in AI allow it to be used to generate many kinds of art in the form of images, music, and even stories. AI-generated arts pose a threat to the livelihood of many artists whose income is reduced due to the decrease in demand. In this paper, we present the result of our study into the different techniques for detecting AI-generated anime images and separating them from human-artist-created images. Using transfer learning, we trained MobileNetV2 and MobileNetV3 models using 750 anime images from a dataset containing 1000 anime images generated using NovelAI and sourced from Danbooru2021 website. We tested the trained models on the other 250 images and our experiment, implemented in Python programming language and using the Keras library, reveals that both models perform well, with accuracy ranging from 96.8% to 97.2%. More importantly, our experiment also shows that both models can retrieve all AI-generated images in the test dataset (100% Precision score) but at the same time incorrectly classify a small number of human-artist-generated images as AI-generated images (Recall score of 94.3% and 95.0%). We argue that, with more work using larger-sized datasets, this approach has the potential to be used in real-world applications to filter out AI-generated anime images from online art marketplaces

Item Type:	Article
Keywords:	Deep generative models, AI-generated images, Anime images, Image classification, MobileNet, NovelAI
Subjects:	000 Computer Science, Information and General Works > 000 Computer Science, Knowledge and Systems > 006 Special Computer Methods > 006.3 Artificial Intelligence, Machine Learning, Pattern Recognition, Data Mining
Divisions:	Faculty of Engineering & Informatics > Information System
Depositing User:	Administrator UMN Library
Date Deposited:	13 Mar 2024 05:57
Last Modified:	12 Jul 2024 02:54
URI:	https://kc.umn.ac.id/id/eprint/29644

Actions (login required)

View Item