Research On Image Retrieval Based On Visual Saliency

With the rapid development of computer technology,it is a very difficult task to find useful information from the large digital image library effectively.Image retrieval is an important research topic in the field of pattern recognition and data mining.In this paper,the image retrieval technology based on visual saliency was analyzed and studied.The specific work was as follows:Firstly,we deeply studied the image retrieval based on visual saliency and its related technologies.The main contents include two aspects.On the one hand,the technology of extracting salient regions included the attention mechanism of human visual,visual saliency model and the detection of interest region.On the other hand,the image retrieval methods included the extraction of low-level visual features,similarity measurement algorithms and the performance evaluation index of image retrieval system.Secondly,this paper proposed an image retrieval method based on color and texture hybrid descriptor(CTHD).Because of the traditional motif descriptors cannot describe the images' texture feature comprehensively.In this paper,five types of novel structural element descriptors are defined to extract image texture feature which completely depicts the case where all texture structures appear in the 2?2 pixel grid.The color features are extracted by quantizing the HSV color space into a single channel with reduced number of 72 bins.The color and textural features are fused together to construct the inherently color and texture hybrid image descriptor which is rotation and scale invariant.This fusion is carried out by extracting textural cues over each color bins independently.CTHD has been tested on the Corel dataset and experimental results suggest that CTHD outperforms other 5 state-of-the-art descriptors.The experiment shows that the color texture retrieval hybrid descriptor method has better retrieval performance.Thirdly,for the traditional image retrieval methods set the same priority for each pixel in the image which cannot simulate the human eye visual perception process.We established the visual saliency model to extracting the salient regions according to human visual attention mechanism.The images' salient region was used for image retrieval.Extracting the brightness,color and direction of the visual features in RGB color space,then,establish five layers of Gaussian pyramids on the three visual features.The saliency map of each feature was extracted according to the principle of concentric circles of receptive field by using central-surround operator.Finally,we combined the three feature saliency map into final saliency map with dynamic weight method.The saliency region of the saliency map was segmented by the method about Otsu algorithm.We compared the method about extracting feature vector on salient region with other methods that extracting the features on the whole image region.The results show that our method performs well.Finally,this paper proposed a novel image retrieval method based on multi-regions color and texture hybrid descriptor.We extracted the underlying visual features of the image in HSV color space.A salient model was established to extract the salient region and background region of the image.In particular,due to only taken the saliency region into account may caused the retrieval process unsteadily,the background region was also used to extract images' feature.The feature vectors of the salient regions and the background regions were extracted respectively,and set the corresponding weights to reduce the mutual influence between these two regions.The multi region image retrieval method was compared with the image retrieval method of extracting only significant area and the other retrieval methods,the results show that the image retrieval method with multi region has better image description intensity,and can representative the content of image well.