Fuzzy Confidence Measures in Midlevel Vision

Segmentation, description, and recognition of objects and regions in natural images involve processes containing varying degrees of uncertainty. High-level vision systems can utilize measures of confidence in scene component labeling to prioritize investigation, resolve conflict, allocate resources, etc., if such measures are provided by the midlevel vision subsystem. A methodology based on the theory of fuzzy sets is presented which can be used to produce linguistic, i.e., natural language, confidence measures of object labeling in natural scenes. The linguistic structure models the uncertainty and complexity and allows for easy incorporation of high-level ancillary information, such as weather conditions and sensor motion. The use of the linguistic confidence structure is demonstrated in two applications. The first involves the labeling of objects in multitemporal forward-looking infrared (FLIR) sequences. The second application is in fusion of information from more than one sensor using sensor models. In both cases, new methods for automatically generating the linguistic values from the image data are described. Also, a new linguistic approximation scheme is developed to generate the final natural language result which reflects the nature of the fuzzy arithmetic.