Occlusion-robust Face Alignment And Recognition Via Locality-constrained Coding

Since twenty-first century,the human society has totally entered a fast-growing information age.With the fast development of science and technology,people have fully enjoyed the convenience brought by the high-tech.Meanwhile,the public are much concerned about their own information security.Under this background,as one of the most popular research topics,the face recognition technology has gained great development in recent years.However,in the past,several studies are conducted in a controlled laboratory situation.How to perform face recognition and achieve not only outstanding but also stable performance in an uncontrolled real scenario along with environment,expression and angle changes is becoming one of the burning issues.Driving by such an urgent demand,based on some hotspots of face recognition,we carry out a series of research work.The main contents are as follows:Considering the face recognition scenes with occlusion but without misalignment,we propose an algorithm called Efficient Locality-constrained Occlusion Coding?ELOC?.During the classification process,we iteratively estimate the face occlusion area at first.On each iteration,the regional mark is modeled by Markov Random Field,and then the areas marked as non-occluded are regularized by local constraint.After several iterations,the query face image is finally identified based on the information provided by the non-occluded area.In this way,we avoid to perform the time-consuming l₁-minimization as well as exhaustive subject-by-subject search.Consequently,our proposed method significantly improves the previous algorithms in efficiency without losing too much accuracy.Moreover,by simplifying the regularization,our proposed method can be further accelerated.Considering the face recognition scenes with occlusion and misalignment,we propose another algorithm called Robust Alignment by Weighted Error and Locality-constrained Representation?RAWELR?.During the alignment and classification process,we firstly estimate the alignment parameters,coding coefficients and error's weights iteratively.On each iteration,we align the query face image by current alignment parameters,use locality-constrained coding to describe the aligned image over the dictionary,and then penalize the reconstruction errors by weighted l₂-norm regularization.After several iterations,the final alignment parameters are applied to align the original face image,and we finish the recognition by utilizing the final error's weights.By doing this,we make fully use of the correlations between different faces.The proposed algorithm attaches more importance to non-occluded pixels when aligning the image.What's more,the interference from those occluded pixels is reduced when deciding the identification of original face image.Therefore,our proposed method still works well even if there are occlusions and misalignment at the same time.In this dissertation,we design multiple experiments for comparison on several public face databases,including Extended Yale B,AR,CMU Multi-PIE and LFW.The experiment results all agree with the feasibility and practicability of our proposed methods.