Low Resolution Face Recognition In Surveillance Video

Biological recognition is a rising technology which utilize human inherent body biological characteristic to identify.Biological recognition technology is better than the traditional recognition method for its security,secrecy and convenience.Face recognition is the most popular research topic in biological recognition due to its un-invasive,concealment and accessibility.Face recognition technology has been widely used in many areas,such as attendance,security system and criminal investigation.Under constraint conditions,the existing algorithm can achieve great results.However,most face images are captured in the non-constrained conditions by the monitoring device.The attendant problems are small size and low quality,which should be named as low-resolution face recognition problems.In this paper,low-resolution face recognition in video surveillance is studied.This paper mainly focuses on face images super-resolution reconstruction,resolution robust feature extraction and classifier training.The major works of this paper are summarized as following:(1)Typical face hallucination algorithm based on patch-wise sparse coding can achieve superior results but has very high complexity in the dictionary learning phase.In order to reduce the time complexity and obtain a relevant dictionary,this paper utilizes K-means++ clustering to train an over-complete dictionary pair.Then,the LARS algorithm is used to calculate the sparse coefficients and reconstruct the high-resolution face images.Experiments shows that proposed algorithm can reconstruct high-resolution faces with rich details quickly in the case of irregular size of surveillance faces.(2)Traditional feature extraction operators have poor robustness at different face resolutions.In order to improve the recognition ability of features,this paper proposed a novel method by combining multi-scale blocking CS-LBP based on Gaussian pyramids and weighted PCA algorithm.Local texture feature are calculated by CS-LBP,and the dimensions of feature are reduce by PCA.Then,the face feature can be extracted for recognition,which has strong classification ability and low-dimension.Experiments shows that proposed algorithm can improve the face recognition accuracy in low-resolution conditions,while reducing the storage space and elapse time.(3)Because the nearest neighbor classifier is easily affected by isolated point and needs to match one by one,this paper studies ELM.In order to build a better model for low-resolution face recognition,ELM classifier is trained by changing the number of hidden layer neurons.In.the further experimental process,this paper develops a recognition platform for the surveillance camera in the laboratory.The recognition results can be achieved by the operation of platform.