Research On Image Set Classification And Its Application

With the rapid development of hardware and software technology,video sequences and multiple static individual images are easier to obtain than ever before.In recent years,image set classification has become one of the research hotspots.As a supervised machine learning technology,it can provide more abundant information of the objects for classification,and it is more robust and discriminant.Therefore,it is widely used in face recognition,object classification,action recognition,scene recognition and other real application scenarios.However,due to the increase of set data,the intra-class variability and inter-class blur of image sets are large,which brings great challenges to image sets classification.Existing image set classification methods have low robustness and accuracies with noise or outliers.Therefore,it is necessary to study image set classification,and make full use of the rich set information to build an effective method for its overall discrimination,so as to improve the robustness and accuracy.For enriching the modeling theory of image set and improving the actual performance of relevant learning methods,it has important theoretical and application value.The main work of this paper focuses on designing robust image set classification methods such that obtain better recognition performance and discriminated ability in actual face recognition applications.In summary,the novelties of this work include:(1)We propose an adaptive correlation learning for image set classification.We use the block diagonalization matrix to weight the correlation between the set data and the representation coefficient,and remove to the redundant balance parameters.Considering the data correlation,an adaptive correlation regular term is proposed to explore the relationship between the probe set and the gallery set,potentially generate sparse discriminant coding coefficients and improve the discriminant and robustness.Theoretically,we proved that the adaptive correlation regularization term can adaptively balance sparse and cooperative learning.A large number of experiments on three datasets prove that ACL has excellent recognition accuracies and robustness.(2)We propose a self-weighted latent sparse discriminant learning for image set classification.To make full use of the encoding information between different gallery sets and improve discrimination of the representation model,it simultaneously minimizes the nearest distance between the probe set and the whole gallery set,and the distances between each independent gallery set and the whole gallery set.To find a more precise nearest point in the gallery sets,it proposes a self-weighted strategy to control the contribution of each gallery set,such that each gallery set can be treated differently to improve its discriminant ability.It proposes a latent sparse normalization with capped simplex constraint to approximate the sparse constraint term for reducing the involved trade-off parameters and computational complexity;meanwhile,it can preserve the robustness of sparse representation.It obtains excellent performance by experimenting on four public benchmark datasets.(3)We propose a partial graph refinement for image set classification.This work is the first attempt to model image set as a graph,which can take full advantage of the set information,label information,and structure information of all images.Different from the existing point-to-point distance criterion,this work proposes a clustering-based classification strategy to mitigate influence of outliers and noisy images.By using a partial graph refining scheme,a high-quality indicator graph can be obtained accordingly.Various experiments show the superiorities of the proposed PGR method in terms of effectiveness and efficiency.