基于深度学习的孟加拉语伪造新闻图片卡视觉取证检测
原标题：Visual Forensic Detection of Fabricated Bangla News Photocards using Deep Learning
AI 摘要
该研究将孟加拉语假新闻图片卡（photocard）检测重新定义为视觉图像伪造问题，构建了包含969张真实和1077张伪造图片卡的数据集，经16种变换扩充至32736个样本。研究对比了全连接ANN（约52%准确率）、紧凑型四块CNN（87.45%±0.53%）和ImageNet预训练YOLOv8n-cls（97.36%±0.42%），并采用无泄漏分层分组五折交叉验证评估。该工作提供了数据集、预处理流程和无OCR的图片卡验证评估协议，为低资源语言虚假信息检测提供了视觉取证新思路。
正文节选
Visual Forensic Detection of Fabricated Bangla News Photocards using Deep Learning Abstract In Bangladesh, a large share of localized misinformation spreads as photocards: single-image news summaries that anyone can fake and that copy the look and branding of well-known news organizations. The trouble is that the misleading message lives in the pixels rather than in machine-readable text, so the natural-language fact-checking tools we normally lean on never even see it. In this work we recast fa