<![CDATA[A delivery note (surat jalan) is a critical document in logistics that demands accurate data recording; however, the manual methods currently employed are often inefficient and error-prone, particularly in high-volume environments. This study aims to evaluate the performance of Tesseract-based Optical Character Recognition (OCR) and Convolutional Recurrent Neural Network (CRNN) in recognizing text on semi-structured documents. Utilizing a comparative experimental approach, this research utilizes a dataset of 200 document images comprising printed text and handwriting under various conditions. A total of 33 images were designated as test data, while the remaining images were used as training data with augmentation. The developed system encompasses image preprocessing, text recognition, and field extraction using regular expressions. Evaluation was conducted using Character Error Rate (CER), Word Error Rate (WER), and Match Error Rate (MER) metrics. The results indicate that Tesseract OCR outperforms at the character level (CER) at 42.84%, whereas OCR+CRNN demonstrates relatively better performance at the word and overall matching levels (WER and MER) at 68.24% and 51.56%, respectively. It is important to note that both values remain very high, a CER of 42.84% indicates that nearly half of all characters are still incorrectly recognized, while a WER of 68.24% means more than two-thirds of words contain errors, rendering the system not yet suitable for practical deployment. However, the performance improvement by CRNN is not yet significant, indicating limitations in the volume and variety of the training data. Furthermore, system performance is highly influenced by document characteristics, where printed text yields better results compared to limited and non-representative handwritten text. In the information extraction phase, structured fields achieve higher accuracy than complex fields, confirming that OCR output quality is the primary factor in extraction success. This study demonstrates that the selection of an OCR method must be tailored to document characteristics and underscores the importance of larger, more diverse datasets to enhance the performance of deep learning-based models.]]>
Copyrights © 2026
