| Author: | Zhao, Mingwei |
| Title: | Car license plate recognition using deep learning |
| Advisors: | Chau, Lap-pui (EEE) |
| Degree: | M.Sc. |
| Year: | 2024 |
| Department: | Department of Electrical and Electronic Engineering |
| Pages: | 50 pages : color illustrations |
| Language: | English |
| Abstract: | This dissertation work introduces the use of attentions in Convolutional Recurrent Neural Networks (CRNN) for vehicle license plate recognition . It aims at improving recognition accuracy without compromising computational efficiency, especially in the real world where conditions could be dynamic and difficult. License plate recognition forms an important constituent of intelligent transportation systems. Yet, practical performance is hampered by poor illuminations, oblique viewing angles, motion blur, and partial occlusions. The research begins with indication of the limitations presented by regular CRNN models. These models poorly perform extraction and sequence modeling of features when faced with complicated and unsteady circumstances. To bridge the above mentioned eventualities, four improvements to these models, each embedding a unique set of attention mechanism, are proposed by this study. First is the spatial domain attention mechanism, which enables the model to focus at the certain most important regions of license plates. Second, self-attention mechanism is incorporated, enhancing the model with very good provision to capture long-range sequence dependencies of input data. Third, a mixture approach of spatial domain and self-attention mechanisms is elaborated: this will not only allow leveraging both localized and global features but also lead to the complementary effects of improved feature extraction and sequence modeling. A dual attention mechanism can be developed and proposed; it integrates channel-wise and spatial-based joint optimizations to maximize recognition performance. Besides improving attention mechanisms, this dissertation looks into the effectiveness of BiLSTM and BiGRU architectures within the recurrent layers of the CRNN. This comparative analysis confirms that BiGRU does not only successfully reduce the computational complexity but also offers a better recognition accuracy overall. The performance of the two attention-enhanced and baseline CRNN models was compared using the CCPD dataset, which consists of a fairly large set of Chinese vehicle license plates. The results showed that the attention-enhanced CRNN achieved a recognition accuracy of 99.86%, which is very good, while being computationally efficient at the same time (in 106.08 seconds for 40,000 plates).This approach strikes a balance between accuracy and speed, making it highly suitable for practical applications. Furthermore, the spatial-domain attention mechanism shows superior efficiency with minimal computational overhead, while self-attention mechanisms exhibit robust performance in modeling long-range dependencies. This research not only contributes to advancing LPR technology but also provides a foundation for further exploration in related domains. Future work will focus on expanding the model's adaptability to diverse license plate formats (e.g., multi-language and international plates), optimizing for deployment on edge devices through model compression, and exploring the integration of multi-modal data to enhance robustness in challenging scenarios. |
| Rights: | All rights reserved |
| Access: | restricted access |
Files in This Item:
| File | Description | Size | Format | |
|---|---|---|---|---|
| 8301.pdf | For All Users (off-campus access for PolyU Staff & Students only) | 1.9 MB | Adobe PDF | View/Open |
Copyright Undertaking
As a bona fide Library user, I declare that:
- I will abide by the rules and legal ordinances governing copyright regarding the use of the Database.
- I will use the Database for the purpose of my research or private study only and not for circulation or further reproduction or any other purpose.
- I agree to indemnify and hold the University harmless from and against any loss, damage, cost, liability or expenses arising from copyright infringement or unauthorized usage.
By downloading any item(s) listed above, you acknowledge that you have read and understood the copyright undertaking as stated above, and agree to be bound by all of its terms.
Please use this identifier to cite or link to this item:
https://theses.lib.polyu.edu.hk/handle/200/13894