Automatic Cropping of Handwritten Scanned Documents with Object Detection Algorithm
Aanchal, Nidhi Nidhi, Preeti Preeti, Gurpratap · Procedia Computer Science · 2023
Automated cropping of handwritten letter fragments with the deep learning-based object detection algorithms is the current trend with an additional technique for the handling of handwritten text recognition with low cost and high speed. Handwriting is utilized to give access to people to do the extra analysis so it is required to be optically scanned as well as a machine familiar method. Owing to unhampered writing ways crosswise within linked and overlying characters, handwriting recognition endures an inspiring job. In this work, the technique is proposed for automatic cropping of handwritten letters from the image of the scanned document using the YOLOv3 algorithm for the detection in the supervision of XML files that contain the corner values of the images which are further learned by the CNN used to generate the bounding boxes. This method is an initial step in surveying the competency of diverse object detection methods in cropping handwritten letters from the scanned credentials. The Punjabi language handwritten dataset has 24 scanned documents used for training and 6 testing images are taken. The training is done on 267 letters and tested on 29 letters. The performance of the method shows the improvement in the accuracy of detection as well as facilitates the process of detection that leading to a reduction in time and costs. The 98.77% accuracy is achieved using the proposed technique which is better than state of art methods.