Comparative Performance and Resource Utilization Analysis of OCR Models for Number Plate Recognition on Raspberry Pi 4

Optical Character Recognition (OCR) is used to capture letters and numbers on license plates, facilitating traffic monitoring, parking management, and toll collection through number plate recognition systems. Since internet access may not always be available, these models must be deployed on embedded devices. Real-time OCR processing on resource-constrained devices poses significant challenges due to limited computational power. Processing OCR in real-time on resource-limited devices is difficult due to restricted computational power. Five OCR models, TrOCR, Doctr, PaddleOCR, EasyOCR, and KerasOCR, were deployed on the Raspberry Pi 4 for the recognition of number plates. Two open-source image datasets were used for performance evaluation. In the pre-processing step, YOLOv8 nano was used for the detection of car number plates. Output was passed to OCR models for the prediction of plate numbers, which were evaluated in terms of accuracy, Character Error Rate (CER), precision, inference time, CPU utilization, and memory consumption. It was seen that Doctr showed an inference time of 378 ms and an accuracy of 90.6%, while TrOCR showed an accuracy of 92.6% but a maximum inference time of 4.98 seconds. Although the inference time for PaddleOCR was lesser at 840 ms, it showed lower accuracy than Doctr. Results conclude that Doctr and PaddleOCR are ideal for the recognition of number plates on edge devices.