Research On Feature Extraction Algorithm Based On Joint Learning

In recent years, pattern recognition is one of hot topics. A recognition model usuallyinclude5parts: data acquisition, preprocessing, feature extraction, feature selection andclassification. Among them feature extraction and feature selection are two key pointsfor classification. They are used to achieve rapid accurate classification, and play a veryimportant role in feature selection and classification. There are many feature extractionand feature selection methods, but in most of the methods these two parts areindependent of each other. These two parts both have their own optimization criterions.However, the intersection of the two optimization criterions maybe not optimum forclassification.This dissertation studies the feature extraction method. Aiming at how to find amethod to joint feature extraction and feature selection together, we improve thetraditional algorithms. The main contributions are as follows:An unsupervised learning algorithm named Joint Sparse Neighborhood PreservingEmbedding (JSNPE) is proposed in this dissertation. This algorithm aims at preservingthe local manifold structure, and achieving row sparsity at the same time. This canreduce the calculation in recognition. It also hopes to reduce the dimension of thesamples, which can make a better result of feature extraction. A good feature extractionresult can also improve the classification result.A supervised learning algorithm named Joint Discriminate Sparse NeighborhoodPreserving Embedding (JDSNPE) is also proposed in this dissertation. This algorithmjoints the discriminate information in the JSNPE algorithm. In this way, we hope topreserve the local manifold structure, and get row sparsity of the transform matrix, andhope to make full use of label information of the samples at the same time. Thealgorithm hopes to increase between-class scatter, and decrease the inner-class scatter.In this way, the algorithm can make samples from the same class closer, and samplesfrom different classes apart from each other.