Areatime Efficient Implementation of Local Adaptive Image Thresholding in Reconfigurable Hardware

Abstract

Local adaptive thresholding plays an important role in image binarization since it is used to effectively distinguish objects of interest from background regions. This step affects the performance of further processing stages in embedded computer vision applications. In local thresholding, a threshold is defined for each pixel as a function of all pixels within a rectangular neighborhood, and as a consequence, this yields a high computational cost requiring significant processing time when thresholding high resolution images or large data sets. This paper presents an area-time efficient hardware implementation of a local adaptive thresholding technique based on the Bernsen algorithm targeted to a field programmable gate array (FPGA) device. Experimental results show that the proposed implementation is resource efficient and able to process a 1024x1024 gray level image in less than 10 milliseconds independent of the neighborhood size. The architecture demonstrates over 100-fold speedup compared to a straightforward software implementation of the original Bernsen algorithm on a desktop computer.