Bangla Document Layout Analysis

This project focuses on developing a Bengali layout detection model to identify and classify elements such as paragraphs, text boxes, images, and tables within Bengali documents. The dataset comprises a diverse range of document types, including historical newspapers, contemporary newspapers, books, magazines, Liberation War documents, and more.

To address this challenge, we utilized the YOLOv8 model, renowned for its effectiveness in object detection tasks. Key strategies included:

• Data Augmentation and Degradation: To enhance model robustness against variations in document quality.
• Model Ensembling: We combined predictions from an ensemble of five YOLOv8 models to further improve reliability.

Our approach achieved Intersection over Union (IoU) score of 0.89637, earning us the 2nd runner-up position in the competition.