Improved Binarization Using Morphology-driven Image Resizing and Decomposition

Abstract

This paper presents a novel binarization algorithm for stained decipherable patterns. First, the input image is downsized, of which the reduction ratio is determined by iteratively applying binary morphological Closing operation. Such morphology-driven image downsizing not only saves the computation time of subsequent processes, but the key features necessary for the successful decoding is preserved. Then, high or low contrast areas are decomposed by applying the grayscale morphological Closing and Opening operators to the downsized image, and subtracting the two resulting output images from each other. If necessary, these areas are further subjected to decomposition to obtain finer separation of high and low regions. Having done the preprocessing, two approaches are proposed to do the binarization: (1) GMM is used to estimate a binarization threshold for each region (2) the binarization problem is treated as an image-translation task and hence a deep learning approach based on the conditional generative adversarial network (cGAN) is trained using the high or low contrast areas as conditional inputs. Our method solves the difficulty of choosing a proper preset sampling mask in conventional adaptive thresholding methods. Extensive experimental results show that the binarization algorithm can efficiently improve the decipher success rate over the other methods.