Rule Based Filtering Approach for Detection and Localization of Bangla Text from Scene Images

摘要

Detection, and localization of Bangla text from natural scene images are important prerequisites for developing Bangla OCR as well as many content-based image analyses. But there is no standard Bangla OCR to be used in the daily work. Due to the presence of some unique features, detection and localization of Bangla text have become more challenging than English text. In this paper, we have proposed MSER based method along with rule-based filtering for efficiently detect and localize Bangla texts from scene images. As the MSER based method is the winning method of the benchmark data, such as ICDAR 2011, this algorithm has been applied to get a better result than related existing methods. By using MSER, candidate text regions are detected. False positives are present into the detected regions. To remove the false positives, rule-based filtering technique has been applied. In this process, geometric properties of text like aspect ratio, eccentricity, Euler number, extent, and solidity have been used to filter out non-text regions. As there is no publicly available database containing scene images of Bangla text, we have developed such database to perform the experiment. The proposed method has been evaluated on 50 sample images of our present dataset containing Bangla and also evaluated on 50 images of ICDAR 2013 benchmark dataset and we have got better results in terms of precision, recall, and f-measure in both cases. A comparison has been made among existing related method and the proposed method and found that the proposed method is better.