One-sample Ear Identification In Unconstrained Environments

The human ear is one of the most common biometric traits that is invulnerable to expressions and aging problem,free from subject's cooperation,and able to be acquired in long distance.Robust ear recognition systems have great application potential in many ways,such as access control,immigration control,law enforcement and criminal investigation.After years of academic research,the ear recognition technology has made great progress.However,when researchers focus on how to improve the recognition rate,some underlying root issues are ignored.Among them,the one sample per person(OSPP)problem is one of the urgent problems to be solved.OSPP problems are often encountered in real-world applications.For example,the police may have only one available image from ID card or monitoring screenshots when tracing a suspect with no criminal records.In essence,the OSPP problem is more than a matter of training sample insufficiency.It is actually the problem that the registered information is inadequate to cover the information of the sample to be identified under certain circumstances,which boils down to the information asymmetry problem.In unconstrained environments,it is equivalent to the partial data problem due to pose variations or occlusions in ear recognition.To address the aforementioned problems,this dissertation carries out a research into the following three parts.1)A Weighted Multikeypoint Descriptor Sparse Representation-based Classification(WMKD-SRC)method is proposed for ear recognition.By adding adaptive weights to all the keypoints on a query image,the intraclass variations are reduced to some extent.Experimental results demonstrate that the proposed method can improve recognition rates of OSPP ear recognition in unconstrained environments,especially in the presence of occlusion and pose variations.Therefore,the robustness of the ear recognition system can be further improved.2)To enlarge the interclass variations among the subjects in the gallery,a method called Weight Optimization of Local Features(WOLF)is proposed.Based on a typical swarm intelligence algorithm,it calculates weight of each local feature from each sample in the gallery.The optimization results are introduced in the above-mentioned recognition scheme.Experimental results demonstrate that the system shows better robustness against pose variations after weight optimization.3)In order to achieve better performance of ear recognition in addressing OSPP problem,this dissertation proposes a 2D and 3D data fusion method on decision-making level and a 2D and 3D data fusion method on feature level,namely,Hybrid Multikeypoint Descriptor Sparse Representation-based Classification(Hybrid MKD-SRC),and Texture and Depth Scale Invariant Feature Transform(TDSIFT).The decision-making-level method works in the sparse representation based framework.2D texture image and 3D depth image are used directly for recognition in the same framework.The feature-level method fuses 2D information and 3D information by encoding a novel local feature descriptor.Experimental results show that these two methods can effectively improve the recognition rate of OSPP ear recognition.Additionally,it reduces the computation time compared with other methods.The study of this dissertation is of great significance in addressing the OSPP problem in ear recognition in uncontrolled scenes,and it is also valuable for other biometrics research in similar cases.The identification method proposed in this study has theoretical significance in solving practical applications,such as secret security,judicial certification,and the police cracking criminal cases.