Research On Key Techniques Of Content-Based Chinese Herbal Medicine Botanic Image Retrieval

This dissertation focuses on Content-based Chinese Herbal Medicine Botanic Image Retrieval (CBHBIR). Based on the rich introduction of some key techniques and the detailed analysis of future trends of Content-based Image Retrieval (CBIR), lots of exploratory research work has been investigated, which includes domain- related feature extraction of Chinese herbal medicine botanic images and hierarchical retrieval method, salient contour extraction and integrated contour matching, regions of interest (ROIs) extraction and matching based on visual attention model, and so on. The presented study is not only the key problems to be settled urgently in the CBHBIR field, but also the current research focus of image processing and information retrieval. Thus, these researches have both the academic and the applied significance.The main contributions of this dissertation are summarized as follows:(1) Several key techniques and algorithms of CBIR are deeply investigated and analyzed, such as, image content descriptors, similarity measures between images, indexical technologies, methods of performance evaluation and relevance feedback, etc. Moreover, the main research directions of CBIR are discussed. And several representative image retrieval systems are addressed.(2) As an important organ of plants, leaf recognition and retrieval plays an important role in plant recognition and retrieval. Because of the inaccurate results retrieved by combining the general visual features, domain-related visual features are analyzed and extracted from the view of plant leaf morphology, such as leaf shape, leaf vein, leaf dent, etc. Owing to the different ability of distinguishing leaves, these features are classified into the global features and the local features. On such a basis, a hierarchical retrieval scheme for leaf images is brought forward. Experiments demonstrate that domain-related features can achieve better retrieval performance, and that the hierarchical retrieval scheme can increase the speed and precision of retrieval system.(3) Chinese herbal medicine leaf images, taken in the natural environments, usually have cluttered background and the leaves partially occluded by other objects, which greatly affects the retrieval efficiency. With inspiration from psychophysical researches of the visual perception of shape, the mechanism of non-classical receptive field inhibition is introduced to contour detection. With such mechanism, the salient boundaries of the main leaves are retained while the short edges in the cluttered background are suppressed. According to the contour detection results, the salient boundaries of the image could be easily extracted, and the image features are thus described by those salient boundaries. Moreover, the integrated boundary matching strategy is adopted to measure the similarity between images. Experiment results show that the proposed methods have excellent retrieval performance.(4) In general, for Chinese herbal medicine flower images, taken in the natural environments, the flower regions usually have prominent characteristic attributes. Motivated by the researches of visual selective attentive mechanism, images are analyzed and the extraction of ROIs is implemented by incorporating the visual attention model proposed by Itti with the seeded region growing. Images are matched by measuring the distances of the features of those ROIs. Experiment results show that the retrieval methods are simple and effective, which can greatly reduce the cost of information processing and increase the efficiency of retrieval system.In a word, we have done some exploratory researches in the CBHBIR field by exploiting the advanced methods and technologies of image retrieval and have proposed some effective approaches which include the extraction of domain-related features and hierarchical retrieval scheme, the extraction of salient boundary and integrated boundary matching, the extraction and matching of ROIs based on visual attention model and so on. The researches presented in the dissertation will greatly accelerate the development of CBHBIR and have both academic and applied significance.