Region-based target identification and multi-feature image retrieval
|School||PLA Information Engineering University|
|Course||Signal and Information Processing|
|Keywords||image retrieval image segmentation object recognition watershed algorithm mean shift analysis of feature space support vector machines evidence theory|
With the advent of powerful but inexpensive computers and storage devices and with the availability of the World Wide Web, automated retrieval of extensive images has. become one of the most popular research focuses. Traditional keywords-based retrieval technology suffers from two main defects such as the vast manual label work and the inevitable subjectivity during the labeling procedure. To overcome the weakness of the traditional method, an extensive and deep study on Content Based Image Retrieval (CBIR) has been developed since 90s of the last century. CBIR implements retrieval using low-level visual features such as color, texture, and simple shape properties. However, CBIR is not yet a commercial success, because most real users searching for images want to specify the semantic class of the scene or the objects it should contain, but there exists a semantic gap between similarity of low-level features and subjective sense of users. At present, semantic image-retrieval gradually attracted more and more attention. Unfortunately, those researchs now remains only exploring with few achievement.This paper studied image retrieval to some extent. The goal is to develop the necessary methodology for region-based multiple feature image retrieval. The work is centred on two generally important issues, namely the image segmentation and the automated recognition of generic object. Some elementary results have been achieved, and an experimental system for image retrieval is implemented,. The work of the paper was evaluated on the experimental system and the advantages have been demonstrated.As for image segmentation, two works have been done in this paper. Firstly, the classical Vincent&Soille watershed algorithm based on immersion simulations is modified. By utilizing the regular spatial information existing in the image grid and a new flooding step, modified algorithm computes a watershed according the basic intuitive definition in a much faster speed than that of the Vincent&Soille’s algorithm do, and most important information useful for segmentation are pereserved at the same time. Secondly, considering the lackness of the feature space analysis-based segmentation methods in region continuity, this paper proposed a mean-shift- and immersion simulation- based image segmentation method. The proposed method adopts a new implementation of mean shift to gain robust estimation of the distinct feature in the image with acceptable time consuming. Also, a new flooding algorithm modified from Vincent&Soille’s watershed algorithm is used to cooperate with the mean-shift results, which take the space information of the image pixels as the most important segmentation consideration.Object recognition is a necessary component of any useful image retrieval system. Since most present successful computer vision object recognition systems can only handle particular objects, some reuseable and robust recognition methods should be developed. In this paper, a new method for image object recognition is proposed. The complicated relation between the visual features and the recognizing result are modeled using evidence theory in the proposed method. Given a recognition task, new method constructs multiple classifying SVMs each for a single feature, and then a modified combination.rule is utilized to fuse initial results from multiple SVMs to a more reliable result as the initial results often conflict with each other. In this way, the influence of different features is tuned properly, thus the system may adapt itself to different recognition tasks. Experiments demonstrate the effectiveness of the proposed method.Based on the works mentioned above, this paper constructed. a region-based multiple features image retrieval system. All works have been evaluated on Amsterdam library of object images (ALOI).