基于生成对抗网络的车牌图像篡改检测数据增广  

Data augmentation for license plate image tampering detection based on generative adversarial network

作　　者：李来源 霍聪聪 谭舜泉[1,2,3,4] LI Laiyuan;HUO Congcong;TAN Shunquan(College of Computer Science and Software Engineering,Shenzhen University,Shenzhen Guangdong 518060,China;Shenzhen Key Laboratory of Media Information Content Security(Shenzhen University),Shenzhen Guangdong 518060,China;Guangdong Key Laboratory of Intelligent Information Processing(Shenzhen University),Shenzhen Guangdong 518060,China;Shenzhen Institute of Artificial Intelligence and Robotics for Society,Shenzhen Guangdong 518060,China)

机构地区：[1]深圳大学计算机与软件学院,广东深圳518060 [2]深圳市媒体信息内容安全重点实验室(深圳大学),广东深圳518060 [3]广东省智能信息处理重点实验室(深圳大学),广东深圳518060 [4]深圳人工智能与机器人研究院,广东深圳518060

出　　处：《计算机应用》2024年第S01期301-308,共8页journal of Computer Applications

基　　金：国家自然科学基金资助项目(62272314)。

摘　　要：现有的篡改检测方法,主要使用基于数据驱动的深度学习模型,检测效果与训练数据的质量和数量成正比,且人工制作高质量的篡改图片费时费力。针对高质量车牌篡改图片数据量少的情况,提出一种针对车牌场景的篡改图片数据增广方法。结合车牌定位模块、车牌矫正模块、基于生成对抗网络(GAN)的图像擦除模块和文字风格迁移模块,构建一个车牌字符篡改系统,以模拟真实场景的车牌篡改流程。相较于传统篡改方法,借助GAN生成的篡改字符种类更多元化、更具备多样性。实验结果表明,使用所提系统生成的车牌篡改图片可以达到篡改区域语义高度合理,且肉眼不可分辨的视觉效果;将它作为扩充数据训练篡改检测模型,曲线下面积(AUC)提升了42.9%,F1值提升了33.0%,漏检率下降了16.6%。同时,使用所提系统生成的车牌篡改图片搭配多种数据处理方法在不同篡改检测网络上均能有效提升检测性能;使用扩充数据训练后,篡改检测网络不仅可以成功检测传统篡改方法的篡改痕迹,针对现阶段流行的生成式篡改,检测效果也明显提升。The existing tampering detection methods primarily rely on data-driven deep learning models,and the detection performance is positively correlated with the quality and quantity of the training data.However,manually creating high-quality tampered images is a time-consuming task.In view of the scarcity of high-quality tampered license plate images,a tampered image data augmentation method specifically for license plate scenes was proposed.By combining the license plate localization module,license plate correction module,image inpainting module and text style transfer module based on Generative Adversarial Network(GAN),a license plate character tampering system was constructed to simulate the tampering process in real scenes.Compared with traditional tampering methods,a more diverse and varied generation of tampered characters were allowed by using the GAN.Experimental results show that the tampered license plate images generated by the proposed system achieves highly reasonable semantic tampering areas and visually indistinguishable effects.By using these images to train tampering detection models,the AUC(Area under ROC curve)is improved by 42.9%,the F1 score is improved by 33.0%,and the false negative rate is reduced by 16.6%.The use of these generated tampered license plate images,combined with various data processing methods,effectively improves the detection performance on different tampering detection networks.Furthermore,after training with augmented data,the tampering detection network not only can successfully detect the tampering traces of traditional methods but also can improve the detection effects of current popular generative tampering methods.

关 键 词：生成对抗网络 图像擦除 文本风格迁移 篡改检测定位 数据增广 

分 类 号：TP399[自动化与计算机技术—计算机应用技术]