Soft Data Fusion for the Industrial Application of Computer Vision

The relevance of information fusion methodologies increases due to the complementary development of computer and sensory technologies. Newly hardware and software facilities allow the inclusion of different information sources in a computer system. Information fusion basically attains the transformation of the information delivered by multiple sources into one representational form. The fused data does not only reflect information that can be extracted from the individual sources but also information not derivable from any of them on its own. Such an information gain characterizes the purpose of information fusion. Operator research in the context of fuzzy systems has generated a fruitful set of aggregation operators, e.g. fuzzy connectives, weighted ranking operators, Ordered Weighted Averaging (OWA) operators, Fuzzy Integrals. So-called fuzzy aggregation operators constitute a flexible alternative to operators traditionally used in information fusion. Among them it is worth pointing out the role of the fuzzy integral. The concept of fuzzy integral is due to Sugeno, who presented in 1974 a mathematical approach within Fuzzy Computing for the simulation of multi-criteria evaluation taking into consideration some cognitive aspects. Sugeno's hypothesis is that the process of multi-criteria integration undertaken by human beings subsumes the linear combination of the different criteria with numerically expressed priorities, i.e. weighted sum strategy. Due to its relationship with cognitive processes and to its positive features as fusion operator, the fuzzy integral is employed in different application fields, where Decision Making and Subjective Evaluation represent the most natural ones. Furthermore fuzzy integrals were used in Computer Vision problems, both on Image Processing and Image Analysis, in a very early stage of research. In this context the fuzzy integral is mainly used because of its mathematical properties as fusion operator, which will be elucidated in the tutorial. In spite of the flexibility, robustness, and interpretability that the fuzzy integral presents when being used as fusion operator, few information fusion applications, especially in Computer Vision, are based on it. This may be due to the complex theoretical background and to the lack of successful implementations of the methodology. Therefore the tutorial brings the fuzzy integral from a mathematical domain to the engineering domain. This goal is achieved in different steps. First, an engineering framework for all fuzzy fusion operators, which is denoted as Soft Data Fusion, is developed. Furthermore different processing frameworks with information fusion, which go beyond the application of the fuzzy integral on its own, are developed. These frameworks are eventually applied for edge detection on color images, for the industrial inspection of high reflective materials, for the processing of document images, the segmentation of color images and for the industrial inspection of end consumer goods. Second the tutorial gives the guidelines underlying the development of different methodologies, which can be employed in the automated parameterization of the fuzzy integral within computational intelligence systems. In this context Soft Computing methodologies present the advantage of being datadriven, what facilitates the implementation of full automated systems for information fusion based on the fuzzy integral. Neurocomputing and Evolutionary Computing are the paradigms selected for the resolution of this problem in the here presented tutorial.