AUTOMATED FEATURE EXTRACTION AND CONTENT-BASED RETRIEVAL OFPATHOLOGY MICROSCOPIC IMAGES USING K-MEANS CLUSTERING AND CODE RUN-LENGTH PROBABILITY DISTRIBUTION

摘要

The dissertation starts with an extensive literature survey on the current issues in content-based image retrieval (CBIR) research, the state-of-the-art theories, methodologies, and implementations, covering topics such as general information retrieval theories, imaging, image feature identification and extraction, feature indexing and multimedia database search, user-system interaction, relevance feedback, and performance evaluation. A general CBIR framework has been proposed with three layers: image document space, feature space, and concept space. The framework emphasizes that while the projection from the image document space to the feature space is algorithmic and unrestricted, the connection between the feature space and the concept space is based on statistics instead of semantics. The scheme favors image features that do not rely on excessive assumptions about image contentAs an attempt to design a new CBIR methodology following the above framework, k-means clustering color quantization is applied to pathology microscopic images, followed by code run-length probability distribution feature extraction. Kulback-Liebler divergence is used as distance measure for feature comparison. For content-based retrieval, the distance between two images is defined as a function of all individual features. The process is highly automated and the system is capable of working effectively across different tissues without human interference. Possible improvements and future directions have been discussed.