Event‐driven relaxation method based on asynchronous parallel processing

Abstract

AbstractThe application of the relaxation method to image processing was proposed by Rosenfeld et al. and has been applied in actual problems. One of the problems in the relaxation method concerns computational complexity, where a uniform processing is applied to all objects in each iterative step. From such a viewpoint, the processing speed should be improved. This paper proposes a region‐growing relaxation method, which is an event‐driven relaxation based on asynchronous parallel processing mechanism, and can be applied to the object for which the interpretation should be modified. From the viewpoint of our transformation model representation, the performance of the relaxation method is analyzed. The relaxation model presented by Rosenfeld corresponds to the fixed‐mean oriented procedure for the object, while the proposed model corresponds to a region‐growing one, where the objects of processing are grouped together. The notion of event and eventdriven processing mechanism is introduced into the pattern information processing, and it is shown that the region‐growing relaxation exhibits a high processing efficiency. The region‐growing relaxation is an extension of the discrete model by Rosenfeld, and has the following features: (i) The convergence (termination) is achieved by a finite number of repetitions; (ii) Asynchronous parallel processing can be adopted; (iii) Algorithm is suited to the eventdriven processing; (iv) The convergence is fast compared with the fixed‐mean oriented relaxation; (v) Region segmentation is achieved at the same time. The region‐growing relaxation method is presently adopted in the newspaper layout understanding system EXPRESS.