Adaptive image segmentation based on visual interactive feedback learning

The advent of cheap, reliable high speed processing has lead to a greater integration of adaptive computing technology into different stages of the manufacturing process, such as the real-time detection and identification of defects on the production line. In this paper we describe the use of an interactive evolutionary algorithm to capture human knowledge in the form of a set of parameters. These parameters control the processing of a set of images from hot-rolled steel surfaces in order to extract "regions of interest"; the resulting segmented images are fed into a defect detection and classification system. In order for this classification system to work correctly it is necessary to extract "regions of interest" with a high degree of accuracy. In a noisy environment, with changing user requirements, there exists a need to easily adapt the parameters to reflect the changing circumstances. We show, that providing the algorithm is suitably tailored to avoid problems of user fatigue, it is possible to evolve optimum parameter sets based on the user's visual evaluation and grading of the resulting segmentations, with corresponding benefits for the manufacturing process.