Deep learning based character recognition method for Ottoman manuscript documents

In this study, a deep learning model for character recognition in Ottoman manuscripts and printed documents is developed and the performance of the developed method for Ottoman character recognition is evaluated with various experimental scenarios. The digitization of Ottoman manuscripts is of great importance for preserving cultural heritage and providing easier access to researchers. However, Ottoman manuscript characters pose significant challenges in character recognition processes due to their complex structure and diversity in writing styles. In this thesis, we focus on the development of appropriate deep learning architectures and various training strategies to overcome these challenges and to achieve high accuracy in character recognition. In this thesis, the Ottoman character recognition problem is considered as an object detection problem and an object detection-based method is developed to solve this problem. Within the scope of the study, two datasets were created, one consisting of Ottoman manuscript documents and the other consisting of Ottoman printed documents. In the creation of the dataset, images obtained from high-resolution scans of the documents were used and character labeling operations were performed on these images. The datasets were used to evaluate the character recognition performance of both manuscript and printed Ottoman documents. The performance of the proposed method is compared with object detection methods such as Faster R-CNN and SSD, which are widely used in the literature. The results show that the proposed method achieves high accuracy rates, especially on printed documents. In this thesis, the effect of different training and testing strategies on model performance is also evaluated. For this purpose, various training scenarios were applied on Ottoman printed and manuscript documents. In the scenario called SC-1, the proposed method performed very well with 98.49% wAP on Ottoman printed documents. In this scenario, the Faster R-CNN model achieved 72.40% wAP, while the SSD model was limited to only 10.64% wAP. This shows that the proposed method is much more effective than the existing methods in terms of character recognition in printed documents. In the SC-2 scenario, the experiments on Ottoman manuscript documents showed that the proposed method performed well with 95.78% wAP. In this scenario, Faster R-CNN performed similar to the proposed method with 96.87% wAP. In the SC-3 scenario, the model is trained using both printed and manuscript datasets and the performance of this model on both datasets is analyzed in detail. The results show that the models trained with combined datasets perform acceptably on both types of documents, but the diversity and complexity of manuscript documents in particular make the character recognition problem more challenging. However, the results obtained in the SC-4 scenario show that the performance of the models trained with only manuscript documents is limited on printed documents. In this scenario, the proposed method achieved 66.41% wAP on printed documents, the Faster R-CNN model achieved 53.12% wAP, while the SSD model achieved only 9.09% wAP. In order to examine the effects of the components of the developed model on the character recognition performance of the model, a series of ablation studies were performed. In the ablation study, firstly, the effect of the backbone structure of the model on the performance of the model was examined. For this purpose, various experiments were performed using different backbone structures such as CSP-DarkNet53 and Tiny-YOLO. CSP-DarkNet53 network is a type of convolutional neural network and has a deeper and more complex architecture. In the experiments, it was observed that the CSP-DarkNet53 network achieved higher wAP values, especially in more challenging character recognition tasks. In contrast, the Tiny-YOLO backbone, despite requiring less computational cost, performed poorly and was particularly inefficient in recognizing complex character structures. These findings suggest that the architecture of a deep learning model is of great importance in the character recognition process. Secondly, the effect of the input size of the model on performance was analyzed in detail. In the study, the input size, which was initially set as 608×1152×3, was reduced to a lower resolution of 320×576×3 and the performance changes were analyzed. Reducing this input size caused a significant decrease in the performance of the model and negatively affected its generalization capacity. This highlights the importance of high-resolution document images in the recognition of Ottoman manuscript characters. In another ablation study conducted within the scope of this thesis, the effect of data augmentation techniques on model performance was analyzed. It was found that data augmentation techniques play an important role in improving the performance of models, especially those running on limited datasets. Data augmentation allows the model to be trained with a wider variety of data, increasing its generalization ability and preventing overlearning. It has also been shown to improve the performance of the model in adapting to the challenging and diverse nature of Ottoman manuscript documents.