The Use of High Resolution Images in Morphological Operator Learning

A critical issue in the design of morphological operators from training data is the limited amount of training images. Recently, a multilevel design approach has been proposed to improve the performance of the designed operators, without increasing the number of training images. Since the operators are usually designed using low-resolution images, this work investigates the use of multiple low resolution images obtained from each high resolution training image as a way of increasing the amount of training data. For the simple down-sampling resolution reduction, this can be achieved using sparse windows without explicitly generating the low resolution images and without any changes in the usual design procedure. Experimental results show that this approach effectively improves resulting operator performance with respect to the mean absolute error for both single and two-level training.