Unified Chinese License Plate Detection and Recognition with High Efficiency (2205.03582v1)

Abstract: Recently, deep learning-based methods have reached an excellent performance on License Plate (LP) detection and recognition tasks. However, it is still challenging to build a robust model for Chinese LPs since there are not enough large and representative datasets. In this work, we propose a new dataset named Chinese Road Plate Dataset (CRPD) that contains multi-objective Chinese LP images as a supplement to the existing public benchmarks. The images are mainly captured with electronic monitoring systems with detailed annotations. To our knowledge, CRPD is the largest public multi-objective Chinese LP dataset with annotations of vertices. With CRPD, a unified detection and recognition network with high efficiency is presented as the baseline. The network is end-to-end trainable with totally real-time inference efficiency (30 fps with 640p). The experiments on several public benchmarks demonstrate that our method has reached competitive performance. The code and dataset will be publicly available at https://github.com/yxgong0/CRPD.

• The paper presents an end-to-end trainable model that unifies detection and recognition of Chinese license plates using innovative deep learning techniques.
• It leverages CRPD, the largest annotated Chinese license plate dataset, to enhance model robustness under diverse conditions.
• The model achieves real-time efficiency at 30 fps and competitive accuracy, supporting applications in traffic monitoring and intelligent transportation systems.

Overview of "Unified Chinese License Plate Detection and Recognition with High Efficiency"

The paper presents an innovative approach to the detection and recognition of Chinese license plates (LP) using deep learning methodologies. The authors introduce the Chinese Road Plate Dataset (CRPD), which, to their knowledge, is the largest public dataset focused on Chinese license plates with detailed annotations. This dataset is designed to address the data scarcity in the domain and acts as a crucial supplement to existing datasets. The central contribution involves a unified network model capable of performing detection and recognition tasks in an end-to-end manner with real-time efficiency.

CRPD: A Novel Dataset

CRPD stands out in its scope as it consists of over 30,000 images sourced primarily from electronic monitoring systems across various provinces in mainland China. The images depict vehicles under diverse conditions, such as different weather, lighting, and states of motion. The dataset contains sub-datasets categorized based on the number of license plates visible in the images: CRPD-single, CRPD-double, and CRPD-multi. Each image is annotated with details such as LP content, the coordinates of the vertices, and the type of the LP, contributing significantly to model robustness across different scenes.

Unified Detection and Recognition Model

The paper introduces an end-to-end trainable model that integrates detection and recognition processes into a single, efficient framework. The detection branch utilizes a learned-anchor-based detection approach, STELA, which reduces computational complexity by associating a single reference box per position, consequently optimizing detection speed. The recognition branch incorporates a framework based on CRNN, composed of convolutional layers, recurrent layers, and a transcription layer, enhanced with deformable convolutions and residual connections for improved accuracy in LP recognition. The use of RRoIAlign allows for precise processing of feature maps, facilitating the effective recognition of rotated text.

Performance and Implications

The proposed model demonstrates competitive performance across existing benchmarks, reaffirming its utility for real-world applications like electronic toll collection and traffic monitoring systems. The real-time processing capability is quantified at 30 fps with 640p resolution, showcasing the balance of efficiency and accuracy. This paper's implications extend to improvements in intelligent transportation systems and inspire future research in extending end-to-end trainable frameworks to handle more complex scenarios in the wild. The release of CRPD is expected to foster more advanced research and development in the field of LP detection and recognition.

Future Directions

This work encourages further exploration of optimizing LP detection and recognition frameworks, particularly by integrating more sophisticated learning techniques or architectures to enhance portability and real-world adaptability. Additionally, the insights gained from this paper may be extrapolated to the field of scene text recognition, aligning with broader efforts to automate and enhance visual data processing systems using deep learning technologies.

In conclusion, the paper presents a substantial contribution towards advancing the field of automatic license plate recognition (ALPR) through the development of the CRPD dataset and a novel, unified neural network model, providing a foundational benchmark and a pathway towards solving data limitations and enhancing model robustness in real-world scenarios.