Research On Image Retrieval And Classification Based On Spatial Relationships

In recent years, the digital images have been springing up rapidly due to the rising popularity of digital cameras, phones with camera and other personal media devices. The development of the second generation internet, interactive community, image sharing platform, social network makes the sharing and spreading becoming so easy. How to efficiently and effect organize these massive images has become a hot topic. The organization and management of images are based on the understanding of image content. Resent years, the most popular topics such as image retrieval, image annotation, image classification and object detection are all based on the extraction of image high level semantic feature.This thesis focues on the representation and extraction of spatial information for image retrieval and image classification and object detection. The main research contents and innovations of this thesis are listed as follows:1. We proposed an algorithm of extraction of image features with contain the object types and locations in images. Then we apply the feature to image retrieval. The traditional image classification based on the spatial relationships use image library which is manually annotated with object type and location. These algorithms are not fit for mass data that appeared recently. In this thesis, algorithms of feature detection, indexing, image matching and ranking are given. The image retrieval system allows user input queries with constant of object relative positional relationships. The experiments show that our approach can deal with the queries comprising explicit spatial relationship more appropriately. It performs better than the existing systems in terms of NDCG@m, MAP and F@m.2. SVMs using spatial pyramid matching (SPM) are popular in recent years. It is been widely used in image classification and object detection due to it prefect classification performance. But they are weak on rotation transformation due to the structure of these pyramids. We proposed an tower matching model to improve pyramid matching model by modifying the hierarchy:using the concentric circles block to replace the rectangle block, using the polar coordinates to replace the rectangular one. The tower is composed by levels, which are divided by three styles:using only radial coordinate, using only angular coordinate, and using whole polar coordinate. Experiments show that the classification results of our method outperform results of state-of-the-art SPMs on Caltech-101and Caltech-256. In scene categorization, our method performs better than ScSPM as well.3.We also give another approach named "Rotation-Invariant Pyramid Matching" to solve the same problem that mentioned above that current SPMs are poor adaptive to diverse viewpoints that images token with. We detect the main directions of objects by edge detection and gradient statistics in images at first, give an smooth rotation algorithm with greedy algorithm and binary optimization to normalize the main directions of objects, build final image features by concatenating features of cells in origin pyramids and rotated pyramids, and finally a linear SVM is employed to classify their images. Experiments show that the classification results of our method outperform results of single kernel image classification approach on Caltech-101and Caltech-256and15-Scenes dataset and this method can combined with other image classification method.4.Traditional supervised object recognition algorithms need manual annotation of the type and location of objects of training dataset, which is less of generalization and is a waste of manpower. In this paper, a unsupervised object detection algorithm is proposed:A LDA topic model is used to analysis, the topics of visual words at first, and an Gaussian mixture model is employed to estimate the size and location of objects. Experiment show that it outperforms existed algorithms.