Project1-AI-Generated-Image-Detection-2026
收藏资源简介:
该数据集是延世大学人工智能系课程“CAS3120·机器学习导论·2026年春季”的项目材料,专门用于AI生成图像检测的二元图像分类任务。数据集包含真实图像和AI生成图像两类,所有图像均为128×128像素的RGB PNG格式。数据规模方面,训练集包含2000张已标注图像,验证集包含1000张已标注图像,测试集包含2000张图像(标签保留用于课程Kaggle竞赛评估)。每个已标注数据集(训练集和验证集)均保持50/50的类别平衡。数据组织上,训练集和验证集通过文件夹名称编码标签:0_real文件夹存放真实图像(标签0),1_fake文件夹存放AI生成图像(标签1);测试集则统一存放在images文件夹中,文件名随机排列且不提供标签。该数据集适用于二元图像分类模型训练、计算机视觉教学实践以及AI生成内容检测研究,特别适合作为机器学习入门课程的教育资源。数据集采用CC BY-NC 4.0许可证,仅限于课程相关的学术和教育用途。
This dataset is project material for the course CAS3120: Introduction to Machine Learning, Spring 2026 in the Department of Artificial Intelligence at Yonsei University, specifically designed for binary image classification tasks in AI-generated image detection. It includes two categories: real images and AI-generated images, all in 128×128 pixel RGB PNG format. In terms of data scale, the training set contains 2000 labeled images, the validation set contains 1000 labeled images, and the test set contains 2000 images (with labels withheld for course Kaggle competition evaluation). Each labeled dataset (training and validation sets) maintains a 50/50 class balance. In terms of data organization, the training and validation sets encode labels via folder names: the 0_real folder stores real images (label 0), and the 1_fake folder stores AI-generated images (label 1); the test set is uniformly stored in the images folder, with file names randomly arranged and no labels provided. This dataset is suitable for binary image classification model training, computer vision teaching practices, and AI-generated content detection research, particularly as an educational resource for introductory machine learning courses. It is licensed under CC BY-NC 4.0, restricted to academic and educational purposes related to the course.
数据集概述
基本信息
- 数据集名称: Project 1 — AI-Generated Image Detection (Yonsei CAS3120, Spring 2026)
- 许可证: CC BY-NC 4.0(仅限学术与教育用途)
- 任务类型: 图像分类(二分类)
- 语言: 英语
- 数据规模: 1K < n < 10K
任务描述
区分真实图片与AI生成图片的二分类图像识别任务。
数据集组成
- 图像尺寸: 128 × 128 像素,RGB格式,PNG文件
- 数据集划分:
train: 2,000张(含标签)val: 1,000张(含标签)test: 2,000张(不含标签)
- 类别平衡: 每个标注集均为50/50平衡分布
文件结构
processed/ ├── train/ │ ├── 0_real/ (标签0:真实图片) │ └── 1_fake/ (标签1:AI生成图片) ├── val/ │ ├── 0_real/ │ └── 1_fake/ └── test/ └── images/ (无标签,图片ID被打乱)
使用方式
通过Hugging Face Hub下载压缩包并解压,示例代码如下: python from huggingface_hub import hf_hub_download import zipfile
zip_path = hf_hub_download( repo_id="legenduck/Project1-AI-Generated-Image-Detection-2026", filename="processed.zip", repo_type="dataset", )
with zipfile.ZipFile(zip_path) as z: z.extractall("/content/")
DATA_DIR = "/content/processed"
提交格式
对测试集进行预测,以CSV格式提交至Kaggle比赛。
id,score 0,0.8234 1,0.1456 ...
id: 对应测试图片文件名(如00000.png对应id 0)score: 实数,数值越高表示越可能为AI生成- 评估指标:AUC(直接基于分数计算,无需设定阈值)
联系方式(助教)
- Donguk Kwon: donguk.kwon@yonsei.ac.kr
- DeogYong Kim: legenduck@yonsei.ac.kr




