Semi - Supervised Image Classification Based On Graph

Supervised learning and unsupervised learning are two traditional learning algorithms in the field of machine learning, which has been widely used in many fields. However, to obtain more labeled samples is more expensive, only using unlabeled data can not get good results in the classification. In response to these problems, semi-supervised learning emerged and has been widely studied and applied. Because of good performance of graph-based Semi-supervised classification algorithm, the objective function is convex, getting the solution is easier and other advantages gained more attention. In this paper, semi-supervised algorithm and expand more in-depth study LapRLS algorithm have been deeply studied. There are several problems in the application of image classification below:highly time complexity and highly space complexity, many samples need to be calculated leads to memory overflow when classify a large-scale image, obtaining a lower classification accuracy rate when the image has complicated background or target.To solve these problems existed in semi-supervised learning based on graph, proposing two algorithms for image classification on the basis of semi-supervised learning based on graph is a good method. The specific works of this paper are as follows:1. The graph-based semi-supervised manifold regularization (LapRLS) classification algorithm needs a large number of unlabeled samples to train a better classifier when the number of labeled samples is small. So it spends much time and space, even out of memory to save the result. If the labeled data is poor, it will obtain a lower classification accuracy rate when the image has a complicated background or target. In order to deal with these problems, a graph-based semi-supervised algorithm combining with mean shift for image classification is proposed. Firstly, mean shift method is used to smooth the image and the result replaces the original image as the image to be classified, effectively eliminates area jump points and noise. Secondly, only a small number of unlabeled samples are used to insteaded of all the unlabeled samples. The experimental results indicate that the proposed method can improve the classifying accuracy and largely reduce the complexity. This algorithm makes it possible for graph-based semi-supervised classification algorithms to classify large-scale images.2. Most graph-based semi-supervised classification algorithms are transductive. It means that predicting and marking new sample data except for the labeled samples and unlabeled samples data is unable. When using transductive graph-based semi-supervised classification algorithms for image classification, because of a large number of image data, the complexity of the algorithm becomes higher, even it can’t compute for the overflow memory, For these problems above, this paper uses the image building plans based on Anchor Graph and effectively extends the scale of solving problems by the transductive semi-supervised graph-based classification algorithms. However, by gathering all sample data together into multi-class with K-means, it will spend a lot of time to cluster into multi-class. Therefore, proposing to use mean shift to obtain the anchor data while avoiding spending a lot of time to cluster into multi-class is a good choice. The mean shift algorithm not only computes anchor, but also smoothes its image as the initial image and improves the classification accuracy. Experimental results show that using the semi-supervised classification algorithms based on anchor can make the transductive graph-based semi-supervised classification algorithms solve image classification problem with a large number of image data and get a better classification.