Small Sample Image Classification Based On Semantic Computation

The small sample problem is the lack of samples and the unbalance distribution of samples in machine learning,which makes the quantity of training data in some classes or total number of training data is small.This widely exists in the image classification field,because the traditional machine learning and pattern recognition method play bad performance on the classes which lack of training data insatiable to actual requirements.The small sample problem is a difficult in the image classification filed.This paper focuses on the small sample problem in image classification,does research on different situations,and finds solutions.This paper takes an application in the image classification field,does research on two different kinds and four different situations with the small sample problems.The four contributions are made in this paper are as follow:(1)In the case of the single source data small sample problem,the problem is a single database small sample problem.Because no other database can be used,the only way is mining knowledge from current database,generalizing models to improve its performance.1)The small sample problem in a single database.In image database,some classes have small number of training data.This is a common problem,especially in a natural scene image since some objects are rare objects,and can supply limited number of training data.In this situation,this paper finds similar data by analyzing the three latent semantic relationship,which are labels of scene,the objects in scenes,the visual words in objects,learn them to the model.A probabilistic latent semantic analysis is proposed in this paper,calculate the semantical similarity,learn the sample under transfer thresholds,to ensure we learn good data,and improve the performance of our method.2)Zero-shot problem,which is the problem that there is no training data for testing classes.This kind of case is also common in physical truth.The database we get,especially the complex scene and natural scene image database,we cannot collect the data of all kinds of objects,annotate them,and treat them as training data.This paper proposes fuzzy attribute by the analysis of fuzzy semantic relationship between local features and attributes,which has better describe ability than its exist binary attribute;finds fuzzy semantic relationship between attributes and classes,proposes fuzzy knowledge,enrich local knowledge by other knowledge sources,and adjust the wrong knowledge.Finally,we propose fuzzy DAP(fuzzy Direct Attribute Prediction)and fuzzy IAP(fuzzy Indirect Attribute Prediction)algorithm.The experimental results show that our algorithms obtain better performance in zero-shot learning than other ones.(2)In the case of the multi-sources data small sample problem,there are other database or database in other nodes can be used,although the semantic calculation,adapt knowledge from other database under specific conditions,improve the generalization of models,increase the classify accuracy.1)In the environment of single machine,the machine store more than one database.Some or one of the database or some classes have a small number of training data,but there are other databases stored in this machine which can be used.This kind of situation usually appears in image database server.In the situation of many databases stored in local machine,some database has little training data,or some classes have little training data.We can use the multi-task learning framework to learn knowledge from other database which stored on a local machine.This kind of other database,even stored in local machine,contain different kinds of data,belong to multi-source data.Multi-task learning learns from multi-source data,improves the generalization by transmitting parameters or models.The model proposed by this paper is different from exist multi-task learning models,we add image understanding in it.Under the thought of image understanding,we take semantic analysis on annotation words of data from different sources,build a semantic binary tree,and merge this semantic binary tree by semantic calculation.The merged semantic binary tree can express the semantic relationship among different classes,guide the learning from other sources data to local data,and increase the classify accuracy.2)The small sample problem in the network environment,there are many nodes in network,which are storing data.The number of training data in the local machine is small,but there exists large number of data in the network.This kind of situation is a common in the network environment now.Although there is limited data in local machine or some class data is limited,other nodes can supply data in network.In the network environment,there are huge data stored in the whole network.The trained classifiers are weak when the number of training data is small in local machine,but the data in network can be used to help training classifier.The difference from former situation,the data ownerships which in network are not belong to local machine,the privacy-preserving must be considered during use these data for training.In this paper,we consider to train weak learners in each node in network,ensemble them to local machine by transfer learning under the calculating weak semantic relationship.This method proposed in this paper consider the privacy-preserving,which the existed transfer learning methods are not considering,and the privacy-preserving is an important element of learning method in network environment.Additional,the method proposed in this paper has the advantage of reducing compute time,reducing compute overhead,reducing communication overhead,and increasing the classify accuracy.In this paper,four different models have been proposed to deal with the small sample problem.The four models design for four different situations of the small sample problem,using different semantic calculation,increase the performance of model in classification,and deal with some specially problem such as privacy preserving in specially situation.