BN-HTRd: A Benchmark Dataset for Document Level Offline Bangla Handwritten Text Recognition (HTR) and Line Segmentation

We introduce a new dataset for offline Handwritten Text Recognition (HTR)from images of Bangla scripts comprising words, lines, and document-levelannotations. The BN-HTRd dataset is based on the BBC Bangla News corpus, meantto act as ground truth texts. These texts were subsequently used to generatethe annotations that were filled out by people with their handwriting. Ourdataset includes 788 images of handwritten pages produced by approximately 150different writers. It can be adopted as a basis for various handwritingclassification tasks such as end-to-end document recognition, word-spotting,word or line segmentation, and so on. We also propose a scheme to segmentBangla handwritten document images into corresponding lines in an unsupervisedmanner. Our line segmentation approach takes care of the variability involvedin different writing styles, accurately segmenting complex handwritten textlines of curvilinear nature. Along with a bunch of pre-processing andmorphological operations, both Hough line and circle transforms were employedto distinguish different linear components. In order to arrange thosecomponents into their corresponding lines, we followed an unsupervisedclustering approach. The average success rate of our segmentation technique is81.57% in terms of FM metrics (similar to F-measure) with a mean AveragePrecision (mAP) of 0.547.