FPGA Parallel Recognition Engine for Handwritten Arabic Words

Abstract

Arabic optical character recognition (OCR) has been an active field of research for decades. Yet, it has limited application domains constrained to printed text. This is partly due to high time complexity of Arabic OCR algorithms. Accelerating these algorithms allows for applications in real-time assistive technologies, office automation, and portable devices. Modern programmable hardware devices like FPGAs have numerous logic resources that allow parallel implementations of many algorithms. In this paper, we investigate implementing the feature extraction and classification stages of handwritten Arabic words on FPGAs. We study the performance and cost of four commonly-used feature extraction techniques and neural network classifiers on images from the IFN/ENIT database of handwritten Arabic words. The most efficient feature extraction technique and the best neural network found are implemented. Multiple FPGA implementations with varying cost and performance are evaluated. An implementation that only consumes about one quarter of the FPGA resources is 20 times faster than the software implementation and is less accurate by only 2.8 %.