AN END-TO-END AUTOMATIC NUMBER PLATE RECOGNITION SYSTEM BASED ON YOLOV12 AND TESSERACT OCR

This study develops and evaluates an end-to-end Automatic Number Plate Recognition (ANPR) system that integrates a YOLOv12-based object detector with the Tesseract OCR engine and controls a physical barrier through an IP-addressable relay. At the pilot site provided by Protel A.Ş., videos in MP4 format were recorded from a barrier-mounted camera under varying daylight, illumination, and color-saturation conditions. The videos were decomposed into frames, and only those containing vehicles were retained. To enlarge the dataset, geometric (horizontal flipping) and photometric (brightness adjustment, Gaussian noise) augmentations were applied, after which licence-plate regions were manually annotated. The trained YOLOv12 model operated within a region of interest (ROI) confined to the lower-central portion of each frame, reducing inference time while preserving accuracy. A preprocessing pipeline, consisting of denoising, grayscaling, adaptive thresholding, and slight blurring, was applied to the detected license plates to enhance character edges, from which alphanumeric strings were extracted using PyTesseract. These strings were then compared against registered license plates in a PostgreSQL database. Upon a match, the backend service sent a command to the relay driver, opening the barrier. In field tests, the system achieved a 96% recognition accuracy over 1,200 real vehicle entries, with an average end-to-end latency of 3.2 seconds per vehicle. The findings indicate that the combination of rapid YOLO-based detection, targeted image preprocessing, and stateless database verification provides a cost-effective ANPR solution for industrial access control scenarios.