Tripartite Architecture License Plate Recognition Based on Transformer

Under natural conditions, license plate recognition is easily affected by factors such as lighting and shooting angles. Given the diverse types of Chinese license plates and the intricate structure of Chinese characters compared to Latin characters, accurate recognition of Chinese license plates poses a significant challenge. To address this issue, we introduce a novel Chinese License Plate Transformer (CLPT). In CLPT, license plate images pass through a Transformer encoder, and the resulting Tokens are divided into four categories via an Auto Token Classify (ATC) mechanism. These categories include province, main, suffix, and noise. The first three categories serve to predict the respective parts of the license plate - the province, main body, and suffix. In our tests, we employed YOLOv8-pose as the license plate detector, which excels in detecting both bounding boxes and key points, aiding in the correction of perspective transformation in distorted license plates. Experimental results on the CCPD, CLPD, and CBLPRD datasets demonstrate the superior performance of our method in recognizing both single-row and double-row license plates. We achieved an accuracy rate of 99.6%, 99.5%, and 89.3% on the CCPD Tilt, Rotate, and Challenge subsets, respectively. In addition, our method attained an accuracy of 87.7% in the CLPD and 99.9% in the CBLPRD, maintaining an impressive 99.5% accuracy even for yellow double-row license plates in the CBLPRD.