Independent Component Analysis Segmentation Algorithm

Abstract

In this paper we propose and investigate a new segmentation algorithm called the ICA (independent component analysis) segmentation algorithm and compare it against other existing overlapping strokes segmentation algorithms. The ICA segmentation algorithm converts the original touching or overlapping word components into a blind source matrix and then calculates the weighted value matrix before the values are re-evaluated using a fast ICA model. The readjusted weighted value matrix is applied to the blind source matrix to separate the word components. The algorithm has been evaluated on 30 overlapped document images from the CEDAR letter database and another 30 degraded historical document images, which containing many different kinds of overlapping and touching words in adjacent lines. Quantitative analysis of the results by measuring text recall, and qualitative assessment of processed document image quality is reported. The ICA segmentation algorithm is demonstrated to be effective at resolving the problem in varying forms of overlapping or touching text lines.