Extracting handwritten characters from the scanned documents is a critical step due to the inherent complexities of various writing styles, inconsistent alignments, multi-touch scenarios, and overwriting characters. Expanding upon the real-time object detection capabilities of YOLOv8 (You Only Look Once), the current paper presents an experiment utilizing a dataset of 2000 handwritten images. This dataset combines the standard dataset (Chars74K) with the custom dataset featuring multi-touch handwritten text, encompassing both individual characters and character combinations that form words. The annotations were created using the Roboflow application and exported to a yaml (yet another markup language) file. The hybrid dataset was split into training, validation, and testing sets. The evaluation process yielded an accuracy of 96.8% at a threshold of 0.5 for recognizing and classifying the characters. The result suggests a positive correlation between training dataset size and model accuracy. Further, fine-tuning the hyperparameters could increase the accuracy upto 98.4%. Additional experiments were conducted to compare YOLOv8 and Detectron2 with Faster R-CNN. The results demonstrated that YOLOv8 offers substantially faster inference times, while Detectron2 with Faster R-CNN exhibited marginally higher accuracy in few classes.