Vertical Margin Detection in Handwritten Manuscripts Using YOLO Object Detection Models

The assessment of handwritten descriptive answer scripts is a critical aspect of educational evaluations, traditionally relying on manual grading, which is time-consuming and sub- ejective. Vertical margins facilitate precise segmentation of answers and identification of question numbers, streamlining the evaluation process. Challenges include inconsistent writing styles, spacing, and varied layouts across different educational institutions that requires a generalizable model. The proposed method encompasses image preprocessing, vertical margin detection, and left- side area segregation using the YOLO model. For tackling the challenge presented by multitudes of horizontal lines intersecting the vertical margin, YOLO v6, v7 and v8 models are used. Dataset is self-collected consisting of varied manuscript consisting of different layouts categorized into different levels – level 1, level 2, level 3 and level 4. The images are captured under different lighting conditions and from multiple smartphones for I inclusion of varied datasets. To provide clarity in the data, histogram equalization is used. Hyperparameters including batch size, learning rate and loss functions are fine tuned to address the challenges posed by lower resolution, perspective distortion, and varying lighting conditions. After evaluation, YOLOv8 model outperformed v6 and v7 with a mAP score of 0.84 and validation as well as training loss in the range 0.15-0.21 which indicates optimal fitting of the model.