Skip to content
This repository was archived by the owner on Dec 23, 2025. It is now read-only.

dxpawn/MobileNetV2-Anime-Detection

Repository files navigation

Enhanced Attention-Based Classification of Human-Created vs AI-Generated Anime Images Using MobileNetV2

[72hr Speedrun] Using deep learning to classify real vs. AI-generated anime images, achieving 97.28% accuracy with attention-enhanced MobileNetV2.
Final term report for the Machine Learning course at VNU-UET, June 2025.


🧠 Abstract

The proliferation of AI-generated anime images presents significant challenges for online platforms, artists, and consumers. In this paper, we address the problem of distinguishing between human-created and AI-generated anime images using deep learning techniques.

We present a novel approach that enhances the MobileNetV2 architecture with channel attention mechanisms and optimized test-time augmentation.

  • Dataset: 5,700 balanced images
    • 2,850 human-created
    • 2,850 AI-generated
  • Accuracy: 97.28%
  • F1-score: 97.29%
  • Architecture: Lightweight, attention-augmented MobileNetV2
  • Strengths:
    • Captures subtle artistic differences
    • Resource-efficient; suitable for deployment
    • Outperforms previous state-of-the-art methods

Our results show that strategic architectural enhancements and inference-time techniques can significantly improve performance without increasing model complexity. This contributes to the growing field of AI-generated content detection and provides tools for protecting human artistic expression.


💥 VeryDeepQuote™

"If a machine can paint a thousand masterpieces in an hour, yet has never felt the anguish of creation or the joy of inspiration, can its output truly be called art? Does the soul manifest not in the final work, but in the struggle to create it?"


📝 Notes

  • No detailed README.md provided — because you can just read the full paper.
  • But hey, thanks for checking this out. ❤️
  • Uploading to arXiv is a hassle, I will try that in the near future -> Got rejected by arXiv (probably cuz it's a course project 😂)
  • If I want to get accepted, I'd probably have to add the "other parts" in + condense paper down to 8 pages - nah, too much work -> Just enjoy the status quo <3
  • Plus I can't find a free anonymous filehoster to host my 7.5GB dataset... so you gotta wait on that (Zenodo doesn't recognize my ORCID, Terminal.LC is closed, I am finding more options)

VGhlcmUncyBhIGhpZGRlbiBtZXNzYWdlIHNvbWV3aGVyZS4uLg== ❓

About

[72hr Speedrun] Using DL to classify real anime images vs. AI-generated anime images, achieving 97.28% accuracy with attention-enhanced MobileNetV2. Submitted as my final term report in VNU-UET's Machine Learning course, June 2025.

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors