复合因素影响下嫌疑人发型变化的深度模拟  

Deep simulation of suspect hairstyles under influence of multiple factors

作　　者：刘耀晖 孙鹏[1,2] 郎宇博 沈喆 孙德廷 宋强 Liu Yaohui;Sun Peng;Lang Yubo;Shen Zhe;Sun Deting;Song Qiang(Public Security Information Technology&Information,Criminal Investigation Police University of China,Shenyang 110854,China;Key Laboratory of Forensic Expertise,Judiciary,Shanghai 200063,China;Civil Aviation College,Shenyang Aerospace University,Shen-yang 110135,China;Criminal Investigation Detachment,Dalian Public Security Bureau,Dalian Liaoning 116000,China;Video Detection Lab,Liaoning Provincial Public Security Dept.,Shenyang 110032,China)

机构地区：[1]中国刑事警察学院公安信息技术与情报学院,沈阳110854 [2]司法部司法鉴定重点实验室,上海200063 [3]沈阳航空航天大学民航学院,沈阳110135 [4]大连市公安局刑事侦查支队,辽宁大连116000 [5]辽宁省公安厅刑事技术总队视频侦查实验室,沈阳110032

出　　处：《计算机应用研究》2025年第3期955-960,共6页Application Research of Computers

基　　金：国家自然科学基金资助项目(61307016);公安部科技计划资助项目(2021YY3);国家级大学生创新创业项目(202110175015);辽宁省研究生教育教学改革研究资助项目(LNYJG2023317);司法部司法鉴定重点实验室开放课题(KF202317);“新时代犯罪治理研究中心”智库项目(20220207);公安学科基础理论研究创新计划资助项目(2024XKGJ0107)。

摘　　要：年龄、伪装等复合因素影响下,命案积案中嫌疑人的相貌、发型等体貌特征变化具有明显的不确定性。针对上述问题,提出双重风格迁移生成对抗网络(dual style transfer generative adversarial network,DstGAN)对人像发型变化进行模拟。首先,设计了双重StyleGAN生成器,借助人像年龄化模型,将人像年龄化信息与发型变化相结合,提高客观因素影响下发型模拟结果的真实度。其次,引入BiSeNET算法对输入人像及其目标发型进行语义分割后得到目标人像语义图,并在FS潜在空间中利用交叉熵损失函数约束GAN逆映射生成的语义图,与模拟后的语义图实现语义对齐,避免出现非自然融合现象。最后,为进一步扩充发型变化种类,通过在RM潜在空间中对发型向量进行编辑,修改输入人像发型所包含的语义属性,实现对于光头等特殊发型的模拟。DstGAN与一些经典发型变化模型相比,更加有效地保证了人脸身份特征的一致性,更加平滑地实现发型与面部边缘的过渡。同时DstGAN在PSNR、SSIM等指标评价的结果中,相比于经典发型变化模型,均取得最为优异的客观评分,表明DstGAN模拟发型变化的人像清晰度更高、感知质量更优、皮肤纹理更真实。The age,disguise,and other combined factors significantly affect the uncertainty of the appearance,hairstyle,and other physical characteristics of suspects in unsolved murder cases.To address this problem,this paper proposed a dual style transfer generative adversarial network(DstGAN)to simulate changes in human facial hairstyles.Firstly,it designed a dual StyleGAN generator,leveraging a facial aging model to combine aging information with hairstyle changes,thereby enhancing the realism of simulated hairstyles under the influence of objective factors.Secondly,it introduced the BiSeNET algorithm to perform semantic segmentation on the input image and its target hairstyle,obtaining a semantic map of the target image.In the FS latent space,it employed the cross-entropy loss function to constrain the semantic map generated by the GAN inverse mapping to align with the simulated semantic map,preventing unnatural fusion.Finally,to further expand the types of hairstyle changes,it edited the hairstyle vector in the RM latent space by modifying the semantic attributes contained in the input hairstyle,achieving the simulation of special hairstyles such as bald heads.Compared to some classical hairstyle change models,DstGAN more effectively ensured the consistency of facial identity features and achieved a smoother transition between the hairstyle and facial edges.Additionally,DstGAN achieves the most outstanding objective scores in PSNR,SSIM,and other indicator evaluations compared to classical hairstyle change models,indicating that DstGAN produces simulated images with higher image clarity,better perceptual quality,and more realistic skin textures.

关 键 词：命案积案 发型变化 风格迁移 BiSeNET算法 语义编辑 

分 类 号：TP391.9[自动化与计算机技术—计算机应用技术]