Mining image content and visual information. Theory and applications

Colantonio S.

Image Mining  Image Content Representation  Image Analysis  Image Understanding 

Mining image content means to extract image hidden patterns, identify image data relationships and, thus, gather novel meaningful knowledge pertinent to the specific domain images belong to. Research in the field is still in its early stages, although it relies on rather assessed disciplines such as Computer Vision, Image Processing, Image Retrieval, Data Mining, Machine Learning, and Artificial Intelligence. The key importance that nowadays characterizes image-based tasks, i.e., tasks that relies on the management, analysis and interpretation of image content, is plainly perceivable in almost all the strategic social, scientific and industrial fields: an imaging investigation is a fundamental step of the medical diagnosis processes; in situ images are acquired for industrial inspection; biometric images are used in surveillance or forensic sciences; georeferenced imagery are gathered and employed in fields such as aerospace, defence, geophysics, intelligence, oceanography, and so forth. Advances in image acquisition and management technologies have fuelled the rapid growth of large and rich image collections, which can reveal meaningful information if suitably processed and exploited. Research in mining image content is just devoted to reach this goal. Image Mining can be seen as the summa and advancement of several processing procedures that are usually applied in image analysis. It requires a long chain that starts with image acquisition and storage, evolves through image processing, image content extraction and suitable representation, image retrieval and indexing, and ends up with the identification of meaningful patterns, thus allowing the production of novel knowledge relevant to the task to be solved. The fundamental challenge in Image Mining is to determine how low-level information contained in a raw image or image sequence can be processed to identify high-level information, and relationships among imagery data, as well as with other contextual data. This dissertation reports the investigation that was carried out in the field of Image Mining by facing several issues related to the different steps of the chain. Theoretical investigations were grounded into the development of innovative methods for tackling all the phases of the image mining process, from image content extraction, representation and browsing, to data mining for the generation of novel knowledge. Such methods were finally integrated into a framework able to support the main image mining functionalities, ranging from image storage to novel knowledge discovery. In accordance with the great value inherent in clinical images, and the increasing amount of digital images available in medical research, medical imaging was selected among the eligible application domains, and some case studies belonging to cardiology and microscopy were considered. More in detail, a novel two-step segmentation method was defined for extracting image structures. A methodological standardization of the features extraction process was proposed by developing a precise classification of the features and an ontological model of the domain. For image content interpretation, a knowledge-based system was developed for solving different types of image-based tasks, by integrating adaptive learning methods, image processing algorithms, inferential reasoning, and meta-level knowledge for strategic planning. Finally, all these methods were integrated into a general framework able to support the image mining chain.