Mahira Jalisha

Built a pipeline that takes in pdf/images of Bengali handwritten texts and gives a text file with digitized version of the text as output. Collected and annotated a dataset with 303 documents. Implemented image pre-processing, line segmentation, removal of overlapping lines, word detection and finally word recognition, Achieved a recognition accuracy of 76%.

Proposed a loss function that improves the segmentation result. Images are in NifTI format. Conducted multiple experiments, each using various techniques and different loss functions. Model used: 3D Unet

I scraped the population data and covid data for each region from the web. I also fixed issues with the geojson map of Bangladesh that was available. The data is available in my github page.