Image Retrieval Method Based On Geometric Distribution Features

With the rapid development of technology and electronic digital devices have worldwide universal accessed to all over the world, so the source of the image data is more and more widely. Every day, there is a large amount of images information spreading and interactive in the network. It has become an important problem that how to quickly and accurately find the information needed by the users in the mass of data and it should be solved as soon as possible. There are two main ways for image retrieval: The traditional text-based image retrieval and content-based image retrieval methods.Text-based image retrieval methods usually adapt the label texts for image retrieval. However, due to the labeled subjectivity, the difference of text description ability and workload effort problems, this method has many limitations. In order to overcome these drawbacks, content-based image retrieval technology emerges. Content-based image retrieval extract representative features from the image content directly, and through the similarity between the image and the target image for retrieval. In this paper, we put forward three basic characteristics of the image based on the geometric distribution for retrieval. Then according to the performance of each characteristic, a comprehensive feature for image retrieval is constructed.In the first step, a concept called main area of the image is referred, and experiments are conducted based on the feature of the width to length ratio of the main area. Firstly separate target object and its background through the image segmentation algorithm, and get the overall outline of the image using edge detection algorithm. Then the connectivity test is utilized to remove background interference which caused by the image segmentation algorithm. The remaining pixels are regarded as the main area of the image and annotated out with a rectangular. The ratio of the rectangular width and length is the characteristic of the image.The second step, we use of representative points to represent image distribution information, and through the relationship between the representative points which are extracted from sub-blocks for retrieval. Firstly, the image is divided into 5 blocks, and the relative entropy of the pixel in sub block area with respect to the pixel in the original image is as a threshold in the process of partitioning, so that the sub-blocks can contain the original image information as much as possible. As the representative points can represent the distribution characteristic in each sub-block area, so can use them to construct an invariant characteristic which can also describe the relationship between different parts of image.The third step is the use of circular ring projection method, which reduces the 2D image into 1D feature vector. Use the geometric center of rectangular image as the same center of the circles which have several certain radiuses. So the original image has been divided into several ring parts. Then account the amount of pixels in each ring region to consist the image characteristic.Finally, we propose a comprehensive characteristic that combine those three features with different weight according to the performance of them. Those features are extracted from the structure of the image distribution. The ratio of the main area width and length only considers the overall distribution characteristics of pixels in the image; The relationship between representative points of block areas feature also take into account the spatial location relationship between the different parts of the image deeply into consideration; The image pixel distribution in ring areas feature is different from the above two features. Those three characteristics are not only independent of each other, but also complement. Experimental results show that comprehensive characteristic method can get good retrieval performance.