Research On Occluded Face Image Recognition Under Natural Scenes

Occlusion face recognition is one of the challenges in real-world face recognition system under natural scenes.Multi-factor changes in the real-world face images may affect the performance of occluded face recognition algorithms,such as resolution,lighting,pose,and expression.The difficulties of occluded face recognition under natural scenes are mainly: on the one hand,partial occlusion caused incomplete face information acquisition and small sample discrimination;on the other hand,diversified occlusion types,uncertain occlusion areas,and random occlusion positions have a huge impact on the accuracy of face recognition.Therefore,about occlusion face recognition under natural scenes,this paper proposes two novel algorithms to improve the recognition rate.The main work of this paper is as follows:(1)Aiming at occluded real-world face images across illumination,pose,expression,and resolution variations,the robust feature is combined with a robust classifier to propose a novel occluded face recognition method.The proposed method is based on gradient feature and intra-class constraint sparse representation classification.First,the texture and contour of face images are extracted through the second-order gradient feature,which is more robust to face occlusion.Compared with the original pixels of real-world face image,this method considers the relationship between adjacent pixels and can reveal the inherent structure of the face.Then,the intra-class constraint sparse representation is presented to reduce the distances between the test images and the specific classes in the gradient domain.The essence of intra-class constraint is to increase inter-class distances and reduce intra-class variations.Furthermore,the intra-class error is used as a final classification criterion.Extensive experiments on different face databases prove that the proposed algorithm can further improve the recognition performance of occluded face images under natural scenes.(2)In order to solve the problem of shadow occlusion caused by light changes on the face image,a light-robust image super-resolution algorithm is proposed.Combining the input real-world low-resolution face image and the face image training set,the low-resolution face images under multiple illuminations are generated by the diagonal loading redundant transformation in the low resolution space,and then the high-resolution face images under multiple illuminations are reconstructed by geometry and position constraints.This algorithm can not only reconstruct low-resolution face images under good lighting conditions from face images with shadow occlusion in the low resolution space,but also can reconstruct high-quality high-resolution face images under good lighting conditions.The algorithm simultaneously reduces the impact of low resolution and occluded shadows on face recognition.(3)This paper mainly analyzes the research background,research content and existing problems of occlusion face recognition under natural scenes from the perspective of theoretical results and practical applications.Secondly,the development status of existing occluded face recognition algorithms are introduced in detail from three aspects: robust feature extraction,robust classifier and deep learning,and their advantages and disadvantages are analyzed and compared.Finally,the theoretical basis and implementation process of classical feature extraction and classification methods are introduced,including image gradient features,sparse representation classification,and collaborative representation classification algorithms,which provide a good theoretical basis for the proposed method.