Data fusion using neuro-fuzzy embedded evidential reasoning

Data fusion provides the power of managing various types of uncertainty (e.g., randomness, imprecision and fuzziness) contained in information while taking into account various types of relationships (e.g. redundancy, complementariness, competitiveness, cooperation and conflict) between information sources. Data fusion has been recognized as an effective and efficient technique which is able to solve complex problems and achieve "greater" quality performance (e.g., better accuracies and robustness) than the use of single-technique based approaches or individual information sources. Although having benefited from many popular theories and techniques, data fusion still presents a quite hard and challenging research area which is calling for new techniques, strategies and theories. Systematically studying hybrid data fusion methodologies (such as meta fusion approaches) is increasingly expected and seems more important than ever in the present information era, where a huge volume of variant types of data of different nature (e.g., meta data) have been emerging with the coexistence of variant types of uncertainty.; Motivated by this challenge, research is conducted in the direction of developing effective hybrid data fusion schemes by comprehensively utilizing the probability theory, the Dempster-Shafer evidence theory, the fuzzy set theory and the neural network technique. Three evidential reasoning data fusion schemes are proposed. They are the neuro-embedded evidential reasoning scheme, the adaptive fuzzy evidential reasoning scheme, and the robust fuzzy evidential reasoning nearest neighbor algorithm. A novel fuzzy evidence structure model is developed based on the availability of probabilistic evidence and fuzzy evidence, for facilitating the application of the fuzzy evidential reasoning to real-world problems. Neural networks are introduced into the Dempster-Shafer evidential reasoning for dealing with dependence between information sources, attempting to relax the required independence assumption by the Dempster's combination rule. A learning scheme is developed. With the proposed fuzzy evidence structure model, the adaptive fuzzy evidential reasoning scheme involves the local discounting and the global discounting for achieving the effective data fusion performance. Furthermore, the fuzzy evidential reasoning cooperates with the nearest neighbor algorithm, leading to the robust fuzzy evidential reasoning nearest neighbor algorithm which can perform the automated selection and combination of nearest neighbors. In the designs of these schemes, research has been carried out over relevant dimensions, including how to model evidence structures, how to measure quality of information sources, and how to adaptively combine fuzzy evidence.; The performances of these schemes are assessed by testing them on two typical data fusion problems: multi-modality magnetic resonance image (MRI) based brain tissue segmentation (classification) and multi-channel remote sensing image classification. Extensive comparisons with other popular methods have shown that the developed data fusion schemes produce better results, in terms of variant evaluation measures, such as classification accuracies, robustness, etc. With further work these schemes can be applied to other application areas, such as data mining, automated target recognition and multi-agent systems, etc.