The Research On Image Retrieval Based On Spatial Structure Information

Images are visual descriptions of objects and are most important media formats. By means of the information in images, people can perceive the external world vividly. With the development of computer technology, the number of digital images has increased explosively. It has become more difficult to find proper images to meet the needs of users in huge databases, so image retrieval is proposed. How to describe image features effectively to find appropriate image feature representation is the key issue in the field of image retrieval. Recently, many researchers have designed lots of methods, but semantic gap between low-level features and high-level concepts has become the limitation of image retrieval. This paper is based on spatial structure information, and takes human visual perception mechanism into consideration. Three image feature descriptors are proposed in this paper. The detail results of this research are as follows:(1)Spatial structure information can describe image details effectively, and is one of the most important characteristics of image. This paper presents an image retrieval method based on color and edge orientation. This method combines color information and edge orientation information together, and calculates the values of edge orientation with edge detection. Based on the analysis of two features, taking into account the relationship between central pixel and neighboring pixels, characteristic values of all pixels are calculated. The numbers of pixels having the same characteristic values are saved to form a feature vector. By using spatial structure information, different position relations between pixels are partitioned in detail. This method can effectively describe color distributions and spatial characteristics, and improve the precision and recall indices for image retrieval.(2)From the microscopic view, an image is composed of many local structures. Local structure can represent image spatial information and be used to analyze image content. It is pointed out that human visual system is very sensitive to the angle changes. Based on the theory,this paper presents a new image feature representation method, namely angle structure descriptor(ASD). ASD is defined by three different angle structures. Based on color information, ASD uses angle structures to detect the image, and extracts the relationship between different pixels in angle structures to establish a feature vector. To a certain extent, ASD can imitate human visual system and describe various features effectively, such as color, texture, shape and spatial information. Experimental results demonstrate that ASD has good discrimination power, and can improve the effect of image retrieval.(3)Biologists have pointed out that human visual perception mechanism includes two stages, visual attention stage and visual perception stage. At the visual attention stage, a lot of visual information is selected for feature extraction. At the visual perception stage, stimulants based on visual information are selected to reconstruct objects in our brains. Inspired by the theory, this paper presents a new image feature representation method, namely multi-trend structure descriptor(MTSD). MTSD is based on three trends in local structure and can explore the internal correlations in images. In addition, MTSD uses three trends to detect color, edge orientation and intensity map, which conforms to human visual perception mechanism.Experimental results demonstrate that MTSD can describe image details and improve the discrimination power for image retrieval.