Information Fusion-based Methods For Image Understanding

The essential task of image understanding is to interpret the acquired image scene accurately. Since image data are some extent fuzzy, the processing of the uncertainty information is of vital importance to all three major steps of image understanding: information acquisition, data representation, and object recognition. Information fusion techniques involve how to process and synthesize information from multi-source, and make them complementary to each other, to obtain knowledge on the object observed which are more objective, more essential than from single source. It is effective to apply information fusion, which is one of the most important fields in intelligent information processing, to process the uncertainty in image understanding. It can be seen as a novel idea which is of high theoretic value and wide application.In this thesis, based on the theories and methodologies of information fusion technology, information acquisition, data representation, and object recognition in image understanding are studied by using pixel-level fusion, feature-level fusion and decision-level fusion respectively.This thesis includes the following contents:(1) The effectiveness of the new approach to image understanding using information fusion are investigated, by analyzing the state of the art of information acquisition, representation of data and knowledge, and object recognition in image understanding.(2) A new method for image data acquisition via image fusion is proposed. Based on the methods of pixel-level fusion and the analysis of the measures of fused image quality in existence, a new evaluation of fused image is proposed. A new method of colorizing night-vision images is also studied, and experiments demonstrate the effectiveness of this method.(3) D-S evidence theory which is often used in uncertainty processing is introduced to the feature-level fusion and its application to the representation of data and knowledge in image understanding are investigated. Key problems and theirs solutions in D-S evidence theory are discussed in details, based on which, new methods for the segmentation of fused image and edge detection are proposed. Experiments demonstrate the excellent performance of the fused algorithms.(4) Multi-classifier-based object recognition using decision-level fusion is studied. A multi-feature hierarchical recognition method based on D-S evidence theory is then proposed to recognize multi-class objects in image understanding. The relationships between D-S evidence theory and fuzzy set theory are investigated and applied in multibimetrics and traffic signs recognition.