Image Semantic Retrieval Based On Salient Region

The rapid development of multimedia technology and Internet technology makes digitalimage resources increase explosively. With the increasing resources promotion universalaccess to information and social development, they provide more information to people thanthey want accept, which makes it too difficult to find the resources they are searching for. Thisdisordered development presents a new bottleneck to both the fast retrieval of numerousimage data and the browse technology. Users are more likely to pay attention to whetherimage regions having certain semantics are similar instead of whether the whole images aresimilar. In order to realize the fast and precise retrieval of image data, this paper poses a newmethod to deal with image retrieval, which utilizes salient regions of image. The main re-search achievements of this paper are as follows:(1) Current methods of saliency detection often generate saliency maps that have lowresolution or poorly defined borders, especially if the salient regions are big, or if the definedborders are poor, This paper introduced a method for salient region detection based on theadaptive spatial neighborhood model(ASNM). Firstly, the proposed method made decom-position of the image gaussian pyramid, which eventually generated different scales of ima-ges. Secondly, differences in form of image brightness and color vision features betweenevery pixel of each image and its neighbor scales were calculated. Finally, the weight matrixwas used to compute every visual feature saliency map. Then the linear interpolation methodwas utilized to get the final saliency map by superimposing feature saliency maps withdifferent scales. After that, the threshold segmentation was made to gain salient regions ofimage. The experiment results show that the recall ratio, precision ratio and Fα of salientregion, obtained by this method, are61.55%、85.44%and70.18%, respectively. The recallratio of this method is4.17%,23.53%,3.73%higher than Itti, Hou and FSRD, respectively.The method mentioned in this paper retains the advantages of accuracy, while at the sametime overcoming the drawbacks of existing methods.(2) In order to improve the ability of computer vision to satisfy the people’s need forunderstanding of visual information and narrow the "semantic gap" between users and com-puter vision, the Probabilistic Latent Semantic Analysis (PLSA) was adopted to obtain imagehigh-level semantic features. The low-level image features upon salient regions, such as color and texture, were firstly extracted. Then the K-means clustering method was used to classifythe obtained low-level image features into K kinds of visual vocabularies, which eventuallyhelped produce the vocabulary description of images. After that, the unsupervised PLSA wasutilized to do image high-level semantics excavation upon the set composed of all imageregions, which finally generating image high-level semantic features. We can obtain that whenthe number of the vocabulary is600, the average classification accuracy is good (87.25%)through experimental analysis. This solves the “semantic gap” between high-level semanticsand low-level semantic, and this provides service for image semantic retrieval.(3) In order to realize image semantic retrieval, OAO-SVM model was used to retrievethe image. The high-level semantic features of training image were regarded as model input,and then SVM model was obtained through learning. In order to improve retrieval perfor-mance, the performance of classification affected by the linear nuclear OAO-SVM suppressorvariable c value has been studied, the result shows that when c=0.7the performance ofclassification is very good, Recall is89.50%, Precision is89.33%, and AUC is0.9401. Withthe test image set classifying model retrieval experiments, it shows that image retrieval resultsbased on the marked area is better than that of image retrieval based on global features.