A Successive Approach to Enhancement of Infrared Facial Images

As it becoming more and more reliable and mature, the technology of face recognition has been widely applied to nowadays life. However, conventionally, face recognition uses the visible light as the imaging spectrum and is thus limited under lighting conditions (such as nighttime) and bad atmospheric conditions (such as rain, fog). Infrared face imaging provides an alternative solution to above problem. Nonetheless, infrared (IR) facial images are usually in low quality due to limitation of current imaging devices as well as atmospheric noises and disturbance. This situation restrains the face recognition system from performing well. Therefore, enhancing of low-quality IR facial images is crucial to a practical IR face recognition system. We in this research work propose to address the problem of IR facial image enhancement by a succession of IR facial denoising and IR facial deblurring. The former is realized via a deep neural network of denoising while the latter is achieved by a blind deconvolution algorithm. The denoising DNN is trained on the Waterloo Exploration database and tested on our multispectral face dataset. The metrics of PSNR and running time are used to compare between different denoising methods including both traditional ones and deep learning-based ones. The metric of Tenegrad is used to evaluate the deblurring method involved. Overall, image quality is improved significantly, which in turn proves our proposed framework of successive IR face enhancement to be beneficial.