Optical Character Recognition Tool for Georgian Handwritten Text Recognition Based on YOLOv8

Optical character recognition is one of the important tasks in the fields of artificial intelligence, computer vision, natural language processing tasks, and machine learning.It remains a major challenge for less common languages such as Georgian.In addition, we face the complexity and variety of writing forms characteristic of this language, caused by the integration of three different scripts implemented over time.Discussed in the presented work are three models: convolutional neural networks (CNN), ResNet, and Yolov8, through which the OCR system was created.The methodology presented in this paper is the first attempt to create for the Georgian language an OCR system based on the Yolov8 platform.The paper analyzes the results obtained by all three methods, based on which we conclude that the accuracy of YOLOv8 compared to CNN and ResNet was improved.After training and evaluating the aforementioned models on the training dataset, it was determined that the YOLOv8 model yielded the best results.The CNN model, incorporating convolutional, pooling, and dropout layers, achieved an accuracy of 0.9230, whereas the ResNet model achieved an accuracy of 0.9302.In contrast, YOLOv8 achieved an accuracy of 0.98848, with a loss value of 0.04479.These results indicate that YOLOv8 performed exceptionally well during live testing.Furthermore, it is noteworthy that during the testing phase, YOLOv8 demonstrated remarkable speed in generating results.