Image Retrieval Based On Qualitative Spatial Reasoning

In recent years, content-based image retrieval (CBIR) has developed into a very activeresearch field. Many internet companies, such as Google, Baidu, Bing and Taobao, haveprovided with content-based image search products due to the20-year’s research. The twokey technologies of designing a CBIR system are image feature extraction and matching. Thestate of the art researches always extract image feature including color, texture and shapefeature, and pay little attention to the spatial relationship of the objects. However the spatialtopological relations between the image objects are not only an important image feature, butalso a reflection of the semantic feature of image, such as “over the sea the moon shinesbright”,“all linked with one another” and “three objects opposite to each other” whichcontain much spatial information. The state of the art spatial relation models in qualitativespatial reasoning field has very few relation predicates, such as RCC-8has only eightrelations. In practical applications of image processing, the relation predicates are fewer andits definition is not rigorous. The descriptive power is weak when the predicates changeaccording to membership. How to construct a powerful spatial model based on qualitativespatial reasoning which can apply to the image retrieval and provide support for feature fusion,the question has not yet been well solved up to now.This paper studies the construction of new spatial models which can solve the actualproblem in CBIR. We focus on studying the spatial topological relations models such as treerepresentation model, layered graph representation model and the topological model betweenconvex regions. We have done the systematic and in-depth study on the inherent defects ofthese methods. We have carried out a series of spatial models according to different situations,which can be fused with shape, density, texture and color feature. These new studies haveprovided very accurate presentation to images and can be applied to image retrieval. The mainresearch contents of this paper are as follows:1. We have completed a comprehensive overview on the state of the art of the imageretrieval problems and made an analysis and discussion of the application prospect andproblems faced in these areas. Moreover, we have made a summary of the field of qualitativespatial reasoning, and introduced kinds of classical spatial relation models. We also havediscussed and analyzed the current situation and existing problems of these models. Theanalysis and discussion of these topics have laid a theoretical foundation for next stage ofresearch.2. Most of the current works describe the binary images with global or statistical features,and they always ignore the spatial information of image objects and other local features. And the single feature description can not accurately represent image. In order to solve suchproblem, we employ a tree representation model (TRM) to describe the topology structure ofmulti-object binary images and fuse several visual features such as shape, density, spatiallocation. In addition, the similar matching algorithm based on TRM is given and applied totrademark database retrieval. Experiment shows our method can represent the image well andis superior to previous methods in search similar binary images with multi-objects.3. At present, the relation models in qualitative spatial reasoning are suitable forrepresenting images with simple objects. The images with complex objects can be representedby internal topology structure and this works are rare. A representative work is layered graphrepresentation model, which only consider the region connected by point as independent partand distinguish from the TRM. So this model can represent image more accurate. However,the layered graph representation model hasn’t considered the tangent relationship between thenode in same level or alternate level. In response to this problem, we are going to extend theabove model and propose a new model named detailed layered graph representation model(DLGRM). We also give similar matching algorithm based on DLGRM and apply it to imageretrieval, the experimental results show that our model can achieve a good accuracy.4. The state of the art topological relation models can be dichotomized to coarse-relationand detailed relation. Most of the existing topological relation models fall into coarse-relationmodels, including classical topological model such as RCC-8. However many applicationsprefer detailed topological relations. To solve the above problem, a new complex topologicalrelation model (CTR) is proposed by traveling the boundary of one region and describingevery single parts of the boundary. We also combined the CTR and shape feature which applyto image retrieval and gain a better accuracy than existing methods.