Distributed Multi-Agent Implementation of Text Line Segmentation Using Parallel Seam Carving

Abstract

Most Optical Character Recognition systems (OCR) and Handwritten Recognition systems (HWR) are based on sequential processing of document images, while some others use parallelism to alleviate the computational load on single processors by balancing the tasks on parallel hardware using “Single Instruction, Multiple Data” (SIMD) architectures. This does not completely correspond to how humans read, where reading a line rely on parallel decision making based on the central and peripheral vision. Thus, it makes these OCR systems limited in face of highly complex OCR problems, unable to synchronize their multiple tasks in a participative way. This paper proposes a parallel text line segmentation approach based on distributed multi-agent systems using a parallel implementation of seam carving. The seam carving text segmentation technique is a language-independent approach capable of processing grayscale images, and showed promising results at extracting handwritten text lines. Additionally, its “top-down” approach to text segmentation offers an opportunity for building efficient parallel OCR systems, using multiple specialized system components built around a multi-agent architecture. This can be further exploited as a base to build more collaborative neuromorphic systems. Tests on different handwritten documents showed an important runtime speedup with parallelization using MPI as a communication interface, without compromising the efficiency.