Multi-task Learning For Face Image Recognition

Face image recognition is one of the most popular research topic in pattern recognition andmachine learning, which is widely used in the information security, public security criminalinvestigation, human-computer interaction and so on. With the increase of the dimension offeatures after feature extraction, feature selection plays an important role in face imagerecognition. On account of the prospection of face image recognition and the importance offeature selection, this paper mainly focuses on single-task feature learning and multi-task featurelearning, studies recognition rate and stability of feature selection of face image recognition. Themain achievements are described as follows:First, we introduce three representative feature selection algorithms such as FS (FisherScore)criterion, reliefF criterion and mRMR (min-Redundancy-Max-Relevance) criterion in detail.Inspired by LPP (Locality Preserving projection), PLS (Preserving the Locality Structure) featureselection criteria is proposed. In the experiment of identification recognition, PLS featureselection algorithm has a better performance on recognition rate and efficiency than otheralgorithm such as FisherScore criterion and so on. From comparative experiments, in face imagerecognition, it preliminarily indicates that the subsets of different feature selection algorithms arenot the same, but share some characteristics or structure among different recognition tasks, whichcoincides with the current research hotspots multi-task learning theory. It is possible formulti-task learning introduced into human face image recognition.Then, face image contains rich information such as race, gender, expression, identificationand so on, corresponding to different information, there is an association among the recognitiontasks. In combination with multi-task learning method, a feature selection algorithm based onmulti-task learning MTLFS is proposed. The experiments not only validate the effectiveness ofMTLFS method, but also show that MTLFS method not only shares the information amongrecognition tasks but also improves the recognition rate and owes a higher robustness. Solvingover-fitting problem of small sample training, MTLFS method is demonstrated the outstandingperformance.Lastly, along with recognition rate and efficiency, stability of feature selection is one of theimportant properties for the evaluation of feature selection algorithms. Based on the relatedresearch in the field of biological information, this paper introduces the concept of featureselection stability into face image recognition. And we give the Measurement method of featureselection stability. Then, we analyze the reasons which influence the stability of feature selection.In several face databases, the proposed feature selection algorithms, PLS and MTLFS methods,show the higher accuracy and stability.