基于生成对抗技术的可见光到近红外人脸图像转换技术  

Visible to Near-Infrared Face Image Conversion Technology Base on Generative Adversarial Nets

作　　者：林志灿[1] LIN Zhican(Practical Teaching Center,Minnan University of Science and Technology,Quanzhou,Fujian 362700)

机构地区：[1]闽南理工学院实践教学中心,福建泉州362700

出　　处：《惠州学院学报》2024年第3期16-21,共6页Journal of Huizhou University

基　　金：福建省教育厅中青年科研项目(JAT190886)。

摘　　要：在光照条件不佳的情况下,红外图像可以更加清晰的显示人脸信息。针对现实中经常出现跨域人脸识别的情况,论文对可见光到近红外跨域的人脸识别技术进行探索和研究,提出了一种改进的基于对抗生成网络的近红外到可见光人脸图像转换技术。对循环对抗生成网络的生成器部分进行改进,增加人脸类别鉴别分支,从而添加人脸鉴别属性约束,使其生成的图像能够更加有效的保留人脸鉴别信息,同时在生成器网络末端引入了非局部算法,使得模型自动关注人脸关键部位,提高跨域人脸识别图像的生成质量。为保证训练得到有效映射,在网络规模较大时,提出了循环一致性损失。实验表明,文中方法在公开的数据集上的性能有所提升。Infrared image can capture face information more clearly in the case of poor light conditions.It is often found that cross domain recognition occurs in practical scenarios,this paper proposed an improved cross domain of visible to near-infrared face image conversion technology based on antagonism generation network.The generator component of the cyclical adversarial generative network has been refined by adding a facial classification branch and incorporating facial discrimination attribute constraints,enabling the generated images to effectively preserve facial identification details.Additionally,the integration of a non-local algorithm allows the model to automatically focus on key facial features,enhancing the quality of cross-domain face generation.To ensure effective mapping during training,a cyclic consistency loss is proposed,particularly for larger networks.Experimental results on public datasets demonstrate improved performance of the proposed approach.

关 键 词：可见光到近红外 循环对抗生成网络 人脸类别鉴别 非局部 

分 类 号：TP399[自动化与计算机技术—计算机应用技术]