基于短波红外线的恶劣环境下跨光谱人脸识别研究

Traditional face recognition techniques based on visible light are susceptible to illumination and constrained to daytime. This problem can be overcome by usage of infrared, which, however, is still bothered by harsh atmospheric conditions such as rain and snow. Short-wave infrared is the solution to all the issues aforementioned. This project studies the problem of cross-spectral face recognition which is advantageous in all-weather scenarios. However, due to the image heterogeneousness issue, neither conventional hand-designed operators for face feature extraction nor the deep neural networks methods can be directly used for the cross-spectral scenario. In this project, we propose a bi-directional heterogeneous image transformation technique which converts the cross-spectral face recognition problem to an intra-spectral face recognition problem. We also design a new deep neural network inspired by the theory of multiple levels of measurement. Furthermore, we consider the issue of quality disparity of heterogeneous images in cross-spectral face recognition and propose a singular value decomposition-based CNN to restore low-quality SWIR images. Coupled with image sharpening and blurring, it deals with this problem via both upward and downward parity, which further improves the recognition performance and realize a robust and effective technique of all-weather face recognition.

基于可见光的传统人脸识别易受光照影响且大多限于白天。采用近红外线可以克服这些缺陷，但仍无法在雨雪等恶劣气候下工作。短波红外线是解决前述诸多问题的有效途径。本项目研究基于短波红外线的跨光谱人脸识别，具有全天候工作的优势。然而由于图像异源性的存在，无论传统手工设计的特征提取算子还是现在的深度神经网络方法都无法直接用于跨光谱人脸识别。因此，本项目提出双向异源图像转换的新思路将跨光谱人脸识别问题转换为单光谱人脸识别，并借鉴多水平测量理论设计新型神经网络进行自动特征提取。此外，项目针对跨光谱人脸识别中的异源图像质量不对等问题，提出一种奇异值分解网络对低质量短波红外线图像进行图像还原，并结合图像锐化和模糊化等方法，从质量向上和向下对等化两方面解决该问题。从而进一步提高跨光谱人脸识别性能，最终实现鲁棒高效且具有全天候能力的人脸识别技术。

基于可见光的传统人脸识别具有受光照变化影响和受恶劣大气条件限制等缺陷，采用红外线波段进行人脸识别可以克服这些问题，但仍无法在雨雪等恶劣气候下工作。短波红外线是解决前述诸多问题的有效途径。因此本项目基于短波红外线成像手段研究具有全天候工作的优势的新型跨光谱人脸识别技术。本研究团队利用深度学习相关工具，我们设计了具有异源特征提取能力的深度神经网络，解决了跨光谱人脸识别的异源特征提取难题，并结合图像质量对等化手段，进一步提升了跨光谱识别算法性能。本研究取得了基于生成对抗的双向光谱人脸转换、多测量水平人脸识别方法、红外人脸检测、红外人脸去模糊等四项核心技术。实验结果如识别率、图像质量指标等均表明，本跨光谱人脸识别技术具有更高效、鲁棒、全天候工作能力等若干优势。本研究工作为跨光谱人脸识别技术提供了新理论和新算法支持，加强了人脸识别的广适性和可靠性，能满足政府、公安、民用安防的户外监控和出入管理等需求，推动了全天候及恶劣条件下人脸识别的应用化进程。