A TWO-STAGE DEEP LEARNING SYSTEM FOR AUTOMATIC NUMBER-PLATE RECOGNITION USING YOLO DETECTION AND A CRNN-CTC RECOGNIZER

Abstract

Automatic number-plate recognition (ANPR) is a key enabling technology for intelligent transportation, used in access control, parking management, tolling, traffic enforcement, and security. This paper presents a two-stage deep learning ANPR system that combines a YOLO object detector for locating the licence plate with a convolutional recurrent neural network (CRNN) trained with the connectionist temporal classification (CTC) loss for reading the plate characters. The first stage localises the plate in the vehicle image as a bounding box, after which the crop is rectified to correct perspective distortion; the second stage reads the entire plate as a sequence, without explicit character segmentation, which avoids the brittle segmentation step that limits classical pipelines. On a licence-plate dataset, the detector achieves a mean Average Precision (mAP@0.5) of 0.97, and the recognizer attains a character-level accuracy of 99.1% and a full-plate (exact-string) accuracy of 96.2% under clean conditions, degrading gracefully to 91.2% full-plate accuracy on a harder subset with blur, skew, and poor illumination. The complete system runs in real time, processing a frame in well under 50 ms on a GPU. Because the CRNN reads variable-length strings, the system adapts to different plate formats with retraining only, making it suitable for deployment on Uzbek and other regional plates. The design complements the authors' broader work on in-vehicle and roadside intelligent-transportation systems.