The Research Of Active Learning Method Based On Sparse Coding In Image Classification

Image classification is an important topic in pattern recognition field.Many image classification system requires user manual annotation ofsample points, and then use the labeled training sample points to train aclassification model. Image data with labels are dificult to obtain, whilethe unlabeled image are easy to obtain.In order to solve the problem ofthe label, active learning has become a hotspot in machine learning andpattern recognition. Now, researchers have done a lot of active learningresearch, such as experimental design and Transductive ExperimentalDesign (TED) algorithm, however, these methods fail to take into accountthe geometrical structure of the data sample point. Based on TEDalgorithm Deng et al.put forward a new kind of active learning algorithmfor text classification, called MAED algorithm, this algorithm for activelearning is performed in the manifold adaptive kernel space,which fullytake into account the intrinsic manifold structure.In this article, by combining the sparse coding of the data withoptimal experiment design, generated a novel active learning methods:sparse adaptive experiment design (SAED). Based on the data set, theSAED algorithm firstly by construct al1-norm graph, and then throughthel1-minimization problem to determine the weights of the graph,finally perform the new weighted graph in experimental design, whichforming a new active learning algorithm. Compared with previous graphconstruction methods, the appliedl1-graph has the following advantages:1.l1-graph does not need the neighborhood size as a parameter. Itadaptively gives the connections between each sample with the remainingsamples.2. The sparse coding ofl1-graph is naturally discriminative which is goodin the view of selecting the important samples.