Based On Significant Regional Extraction And Plsa Of Image Retrieval Method

With the development of multimedia technology and Internet technology,digital image resources are growing at a geometric level,that poses new technology challenges to how to achieve a large number of image data's fast retrieval and browsing. In Content-based Image Retrieval(CBIR),at most cases users do not care about whether the whole images are similar or not,but more concern about the certain semantic area of the image.To compensate the lack of using global features to describe the image content,this paper proposes an image retrieval method based on semantic feature of salient regions.First,we use spectral residual and multi-resolution analysis to detect the salient regions.Then latent semantic model is achieved by using probabilistic Latent Semantic Analysis(pLSA).Finally,region latent semantic feature is obtained by applying pLSA model to each salient region in an image,and this semantic feature can be used to construct a SVM model to fulfill the image retrieval.Based on the spectral residual model,this paper discusses the salient region detection technologies,region latent semantic feature construction technologies,as well as image retrieval based on salient region technologies.The achievements of the research are as below:(1)Salient region detection:According to the characteristics of human visual system,a salient region detection method based on spectral residual and multi-resolution is proposed.We first compute the spectral residual of three features i.e. intensity,color and orientation under different scales to build series of multi-resolution saliency maps,which can be combined through linear interpolation to generate three feature-saliency maps.Then we use k-means clustering for binary clustering and select the feature-saliency map with the largest distance between two centroids.Finally we apply dynamic threshold segmentation to get salient regions in an image. The innovation of this thesis is the salient region detection method.(2)Region latent semantic feature construction:After detecting the salient regions,we carry on latent semantic mining of the set of this image regions at an unsupervised way by applying pLSA,thus to construct the region latent semantic feature.(3)Image retrieval based on salient region detection:Image retrieval can be regarded as a real-time classification if we consider the positive and negative samples as two types. Using Support Vector Machine for learning of the region latent semantic features,and we can get each sample's impact of the decision-making,namely,the training received support vectors,and then use those support vectors to classify the test images,so we can get the final image retrieval results.This paper combines salient region detection method with pLSA to get the region latent semantic feature,which will be applied to image retrieval.This reduces the semantic gap between low-level features and high-level semantics. From the differences of the global based method and our image retrieval method,we could obtain the result that the accuracy of latter image retrieval method is higher than the previous method.