Performance Characterization and Parallelization of Tesseract Optical Character Recognition on Multicore Architectures

Optical Character Recognition, or OCR, is one of the major topics in computer vision technology. It is widely used in various applications, such as a digital libraries, automatic banking systems, and mailing services. Tesseract OCR Engine, which we evaluate in this paper, is one of renowned OCR programs. It was originally developed by Hewlett Packard Lab between 1985 and 1995, and has been maintained by Google since 2006 [1]. Initially, this program was designed to recognize English text only, however, it has been enhanced to support other languages as more training models were added [2]. OCR process including Tesseract is known to be very compute intensive because of the computation involving image and mathematical processing to obtain higher recognition accuracy. While there has been a significant improvement in the recognition accuracy of Tesseract OCR, parallelization has not been extensively studied. Also, there is a plethora of multicore architectures, thus it is highly considered to achieve better performance by utilizing those parallel architectures. In this paper, performance characterization has been performed using a profiling tool to find a target, and then, parallelizing the identified target using an appropriate parallel programming method. The main goal of this paper is characterizing the performance of the Tesseract OCR program and accelerating compute-intensive loops on multicore architectures. Our main contributions are as follows. First, we analyze the Tesseract OCR program using a binary instrumentation tool and identify target loops that can be parallelized. This allows us to decide which parallelization method needs to be applied and how to modify the loops in order to make them run in parallel. Second, we apply parallel methods on the loops based on the characterization analysis. Lastly, we discuss issues and limitations in parallelizing the Tesseract OCR and suggest appropriate solutions as needed.