Document Image Analysis for Deep Learning-Based Text Recognition

Abstract

This study evaluates the performance of two models, OCR + CNN and Hybrid CNN-RNN, using the MNIST dataset, which consists of 60,000 training samples and 10,000 test samples, each sized 28 x 28 pixels. The results show that the Hybrid CNN-RNN model significantly outperforms the OCR + CNN model in terms of Overall Accuracy, F1-Score, Character Error Rate (CER), and Word Error Rate (WER). The Hybrid CNN-RNN achieved an overall accuracy of 99.18%, compared to 93.14% for OCR + CNN, and demonstrated much lower error rates (CER and WER of 0.82%) compared to OCR + CNN (CER and WER of 6.86%). In terms of training and validation accuracy, the Hybrid CNN-RNN also performed better, reaching 99.53% training accuracy and 99.18% validation accuracy at Epoch 5, while OCR + CNN achieved 95.89% and 95.97%, respectively. Despite the superior accuracy, the Hybrid CNN-RNN model required more inference time, taking 7.46 seconds for 10,000 samples, as opposed to 5.22 seconds for the OCR + CNN model. In conclusion, while the Hybrid CNN-RNN model offers better accuracy and stability, the OCR + CNN model is more efficient in terms of inference time, and the choice of model depends on whether higher accuracy or faster inference is prioritized.