Page 125 - Kỷ yếu hội thảo khoa học quốc tế - Ứng dụng công nghệ mới trong công trình xanh , lần thứ 8

108 TRƯỜNG ĐẠI HỌC SƯ PHẠM KỸ THUẬT - ĐẠI HỌC ĐÀ NẴNG

Fig. 1. Sample of a Vietnamese Student ID Card
C. Multil-Lingual Model Development Fig. 3. Model building workflow
Our study demonstrates using BERT and E. Mobile Application Implementation
TensorFlow for OCR model training. Building on this The React Native framework from the initial
approach, customized BERT models will be trained project will be used to develop cross-platform mobile
for each language using transfer learning on the apps on iOS and Android. The interface will guide
corresponding ID card images. Other architectures users to capture ID card photos and display
like multi-headed CNNs will also be explored to recognized text in their native language. Cloud APIs
optimize accuracy and efficiency. The models will be will provide enhanced services when connectivity is
trained to leverage GPUs for accelerated deep available. Figure 4 below shows the UI of the
learning. Figure 2 below shows the flow of text finalized app.
classification with BERT within TensorFlow Lite.

Fig.4. User Interface of the App
F. Testing and Evaluation

Fig. 2. Text Classification with BERT
in TensorFlow Lite
D. On-Device Optimization
To deploy the models on mobile devices, they will
be converted to TensorFlow Lite format and quantized
as detailed in our study. On-device versus cloud-
based inference tradeoffs will be evaluated based on
performance constraints. The TensorFlow Lite Micro
runtime will enable low-latency execution on mobile
hardware. Figure 3 below shows our proposed model
building workflow.

Fig. 5. 100% Accuracy for the card
ISBN: 978-604-80-9122-4