An Intelligent Classification Model for Camera Captured Unstructured Handwritten Documents with Intra Class Similarities

The overall aim of this study is to classify camera-captured handwritten document images with intra-class similarities and unconstrained layouts. We investigate at a dataset of real-time handwritten notes that are frequently utilised for academic assessment in higher education institutions. For the classification of handwritten document images, we introduce a deep learning-based methodology. The models used for integration into the proposed workflow for document classification are ResNet50, MobileNetV2, and Inception ResNetV2. A database of 11676 samples from twelve different classes and three different scripts has been obtained for the purposes of experiments. Depending on the subjects, each class of documents is established by determining keyword-based characteristics. The learning performance of deep learning models is analysed in terms of training, validation, and test sets, which are sampled in the ratio of 80 to 20. The results show that document classification is effective with MobileNetV2, with an accuracy of roughly 99 percent for training and validation datasets and 92 percent for test datasets, with average precision and recall of 86 percent and 82 percent for twelve categories, respectively.