Face Recognition On Low-resolution Images Based On Unified Feature Space

Face recognition is one of the most important branches in biometric identification.It has developed from manual to automatic recognition, from the harshly controlledlab environment to recognition under complex wild conditions. Nowadays, more andmore widely the surveillance networks covered, however, the surveillance camerasusually get poor resolution images because of the limitation of the equipment,technology, cost control and long-distance range. So far, most of the researches inrobust face recognition is aimed at the poses, illuminations, expressions, occlusionsand so on. And low-resolution face recognition is still a relatively new field.Recently, more and more researchers have paid their attention to the lowresolution face recognition, and have proposed some great low-resolution facerecognition algorithms. However, there is no literature review on low-resolution facerecognition. Therefore, for the first time, we summarize the low-resolution facerecognition and classify the low-resolution face recognition algorithms, also weintroduce the kernel theory, the advantages and disadvantages of every proposedalgorithms. We classify the algorithms on the criterion of strategies which solve themismatch of high-resolution and low-resolution feature spaces. And these algorithmsare classified into three groups: Super-resolution based Low-res face recognition,Resolution-robust face representation based Low-res face recognition, and Unifiedfeature space based Low-res face recognition.We propose pose-robust face recognition for low resolution images via unifiedfeature space, and utilize the relations between high-resolution images andlow-resolution images. First, in the training phase, we are looking for a unified featurespace into which features of the high resolution gallery images and low resolutionprobe images are simultaneously transformed, and the distances between themapproximate the distances which the probe images had been taken in the sameconditions as the gallery images as much as possible. Then in the testing phase,transform the high-res and low-res image features into the learned unified featurespace, and classify and recognize directly. Experiments on the CMU MultiPIE datasetindicate that under different poses, illuminations, and resolutions, the proposedmethod get an up to91%recognition rate. And experiments on the challenging, wild MBGC/FRGC dataset further signify the effectiveness of the proposed method, whosecorrect recognition rate up to49%with a0.1pulse acceptance.For the reason of insufficient information which low-resolution images cansupply for face recognition, we introduce sparse-representation-based super-resolution(SRSR) into the proposed pose-robust low-res face recognition method. Therecognition rate can be raised, if the low-resolution images are processed throughsuper-resolution technology, and then recognized by our proposed method. In theCMU MultiPIE dataset, on images with pose changes and a3resolution ratio,recognition rate of the proposed pose-robust low-res face recognition algorithm ishigher than the sparse representation based super-resolution up to10%. Further if getthem together, the correct recognition rate improves around6%, which is up to89%.