Research On Video Face Recognition Based On Image Set And Video Sequence

Focus on the topic of video based face recognition,this paper expounds the research background and significance of video face recognition firstly.And then three types of video face recognition methods which includes video face recognition method based on key frame,based on image set and based on video sequence are reviewed.Based on the related theory and research,in order to improve the recognition rate and recognition performance,several video based face recognition algorithms are researched further.The main contents of this paper includes the following three aspects:(1)The main problem which the existed video based face recognition method based on key frame faced is the definition of key frames is vague,and these methods make no use of contextual information in the video,while time and motion information play a important role in video based face recognition.At the same time,the feature extracted by traditional video based face recognition method of feature is single,lead to low recognition rate,in order to ensure the recognition rate of video based face recognition,common method is to design more complex recognition process,which leads to the increase of the time cost.Aiming at representating face effectively and realizing robust video face recognition with lower computational complexity,this paper proposed a video face representation and recognition method from muti-views which extract LBP codes of three orthogonal planes from Gabor transform domain.Firstly,we get the amplitude map after Gabor wavelet transform with different scales and directions on each face images in video,and then get local texture feature by extracting LBP codes from three orthogonal planes,in the last stage of video face recognition,using weighted Chi square probability Nearest Neighbor method based on Fisher Criterion.Through the relevant experiments which are carried out on the Honda/UCSD video database,verify the validity of the algorithm,and the proposed method is robust to illumination variation and facial expression changes.The innovation of the proposed method is that can realize effectively feature extract from combined time domain,space domain and frequency domain,reach effectively video face representation,and the process of feature weight learning can guarantee the recognition precision of the algorithm.At the same time,the algorithm has lower time complexity.(2)The video based face recognition method based on key frame faced the primary work of how to determine the representative frames of the whole video,but the video target randomly appeared usually,and video environment will be accompanied by changes in illumination,target attitude changes such general factors that interfere with the performance,these noise made the key frame is difficult to locate accurately,resulting in the recognition rate of video face recognition methods based on key frames are generally low.To solve the problems above,this paper proposes a multi-instance learning video face recognition algorithm based on weighted TPLBP.We treate the video face recognition as a multi-instance problem,take each face video as a package,and the normalized face image of the video as an instance,extract cascading weighted blocked TPLBP histogram as feature description.And a multi-instance can be trained through texture feature space of instances in the training set,and then realize the classification and prediction of the test video bag.Through related experiments on the Honda/UCSD database,the algorithm achieves high recognition accuracy,which verifies the validity of the proposed algorithm,and the proposed method is robust to uniform illumination change and pose variations.The innovation of the proposed algorithm is that treate video face recognition problem as a multi-instance problem,effectively avoid the traditional methods need to determine the key frames in the video,and the combination of TPLBP descriptor and multi-instance learning makes the algorithm achieved a high recognition rate,at the same time,it.s robust to the uniform illumination changes and pose variations.(3)Video based face recognition based on image set has been widespread concern in the industry recent years,with respect to a single image,image set can provide more identifying information,can effectively improve the recognition performance of video based face recognition algorithms,these methods involve two main problems which are how to model image set and how to measure the relationship between the models.This paper proposed a video face recognition algorithm based on convex hull of kernel subspace sample selection.The algorithm will treate each video as an image set,extract the feature vector of each face image to represent the face,and the feature vector will be mapped to the kernel space and then carry on sample selection,establish convex hull using the selected sample collection to get the video face model,and the similarity between models can be measured by the distance between the convex hull,a classifier can be obtained through learing from a large amount of the training dataset,and then realize the identification and classification of the test video.Through the experiment on You Tube Face database,verified the recognition performance of the propoed algorithm.The innovation of the proposed algorithm is that through the modeling process of image set,we use the sample selection algorithm to reduce the errors in the reconstruction of samples and other samples,which ensured the selected samples linearly independent,and get effective category information by small number of samples,and the accuracy of the algorithm is improved.To sum up,this paper mainly studied the video face recognition problem and methods,and several video face recognition algorithms are proposed,and some innovative research results have been obtained,and has carried on the forecast to the future development of video based face recognition.