Research On Image Semantic Mapping With Echo State Network

With the development of high-tech,computer science,Internet and storage technology for example,the amount of digital images is growing at a phenomenal rate.So the faster and more accurate methods are need to be created to help people search images.But conventional content-based image retrieval method get similarity between different images only with the distance of low-level features,the retrieval result is not ideal as the problem of semantic gap between low-level features and high level semantic.Therefore the semantic-based image retrieval method is proposed and develop rapidly,it gradually become the research focus in the field of image retrieval.Now,traditional machine learning methods which with the disadvantages of slow training speed,low generalization capability are often used to realize image semantic mapping by researchers.To improve the real-time performance of image semantic mapping,Echo state network are used to be the key algorithm of image semantic mapping method,at the same time,Ensemble learning is used to the model of image semantic mapping to improve the generalization capability of learning machines.Finally the semantic result is used to improve retrieval accuracy.The contents of this thesis are described as follows:(1)The study was to optimize the training process of echo state network classification model.The variation,rate of changes and changing trends of neurons in reservoir are used as the judgment to decide whether or not the training ended.And thus effectively avoid oscillation and divergence in the training process.(2)In order to improve the speed of image semantic mapping,the echo state network classification model which with high training speed are used,it's also hard to be trapped into local optimization.At the same time,the ensemble learning method is used in this thesis.Low-level feature are divided by feature types,after which the divided features are placed into different reservoirs,so that the multi-reservoirs echo state network image semantic mapping model is formed.Ensemble Learning is easily constructed by parallel computing,So the multi-reservoirs echo state network model is designed in the form of multi-core parallel.Experiments result show that the image semantic mapping model proposed with greater accuracy.(3)At the basis of image semantic mapping,the semantic result are use to image retrieval method.The color fuzzy correlogram(CFC)is used as the image features forimage retrieval which is proposed at the basis of color autocorrelation correlogram(CAC).This method not only retains the low space complexity of CAC,but also compute the relations between different color value use fuzzy function.The semantic result computed by semantic mapping model are used to filter the image set to reduce the retrieval scope.the retrieval ability are thus improved.