One important aspect of assessing the quality in\npulp and papermaking is dirt particle counting and classification.\nKnowing the number and types of dirt particles present\nin pulp is useful for detecting problems in the production\nprocess as early as possible and for fixing them. Since manual\nquality control is a time-consuming and laborious task,\nthe problem calls for an automated solution using machine\nvision techniques. However, the ground truth required to train\nan automated system is difficult to ascertain, since all of the\ndirt particles should be manually segmented and classified\nbased on image information. This paper proposes a framework\nfor developing and tuning dirt particle detection and\nclassification systems. To avoid manual annotation, dry pulp\nsheets with a single dirt type in each were exploited to generate\nsemisynthetic images with the ground truth information.\nTo classify the dirt particles, a set of features were com-puted for each image segment. Sequential feature selection\nwas employed to determine a close-to-optimal set of features\nto be used in classification. The framework was tested both\nwith semisynthetically generated images based on real pulp\nsheets and with independent original real pulp sheets without\nany generation. The results of the experiments show that\nthe semisynthetic procedure does not significantly change\nthe properties of images and has little effect on the particle\nsegmentation. The feature selection proved to be important\nwhen the number of dirt classes changes since it allows to\nimprove the classification results. Using the standard classification\nmethods, it is possible to obtain satisfactory results,\nalthough the methods modeling the data, such as the Bayesian\nclassifier using the Gaussian Mixture Model, show better\nperformance.
Loading....