Developing new techniques to improve licence plate detection systems for complicated and low quality vehicle images
Developing new techniques to improve licence plate detection
|Author||Al-Shemarry, Meeras Salman Juwad|
|Institution of Origin||University of Southern Queensland|
|Qualification Name||Doctor of Philosophy|
|Number of Pages||271|
|Digital Object Identifier (DOI)||https://doi.org/10.26192/cbeg-yw22|
Intelligent transportation systems (ITSs) play a very important role in people’s lives in many respects. One of the most important ITS applications is for automatic number plate recognition systems. Over the years, many algorithms have been developed for detecting licence plates (LPs) from vehicle images or from a sequence of images in a video. Many existing ITSs work only under good conditions or normal environments.
It is still challenging to find effective techniques to identify LPs under difficult conditions, such as low/high contrast, bad illumination, foggy, dusty, or distorted by high speed or bad weather. New techniques are needed to improve the performance of existing detection systems.
In this thesis, novel methods are developed for licence plate detection (LPD) systems to extract key features, and classify the LP region from complicated vehicle images based on preprocessing methods and machine learning algorithms with several types of texture descriptors.
In order to identify LPs from complicated vehicles images, four LPD methods were developed in this research. The first, is a three-level local binary pattern operator based on an ensemble of Adaboost cascades classifiers (3L-LBP_Adaboost) detection method. The second method, introduces a new texture descriptor based on a multi-level preprocessing stage with extended local binary pattern descriptor using an extreme learning machine classifier (MLELBP_ELM). The third, develops learning-based preprocessing methods using a local binary pattern and a median filter histogram of the oriented gradient with support vector machine classifier (LBP_MHOG_SVM) for detecting complicated LPs. Finally, for identifying distorted LPs using hybrid features, median robust extended local binary pattern and speededup robust with an extreme learning machine classifier (MRELBP_SURF_ELM). The experimental results show that both of the LBP_MHOG_SVM and MRELBP_SURF_ELM algorithms perform very well in LP detection accuracy rate compared with 3L-LBP_Adaboost and MLELBP_ELM algorithms. Also, the false positive rate (FPR) for both methods is better than those algorithms. The MLELBP_ELM and MRELBP_SURF_ELM methods carry out significant classification of different types of LP key features. The 3L-LBP_Adaboost approach takes much less execution time and produces high FPR compared to the three other methods. But it was a good technique for selecting suitable preprocessing and extraction methods, for detecting LPs from low quality vehicle images.
The experimental results proved the efficiency of the proposed approaches for detecting difficult regions of the LP inside a vehicle image with a high accuracy rate and low detection time. Whereas the overall performance evaluation for the 3L-LBP_Adaboost method in terms of detection, precision, and F-measure rates is 98.56%, 95.9%, and 97.19%, respectively, with an FPR of 5.6%. The average detection time per vehicle image was 2.001miliseconds.
In the MLELBP_ELM method, detection accuracy and FPR were improved by 0.54% and 0.56%, respectively, compared with the 3L-LBP_Adaboost approach. The classification and detection rates are 99.78% and 99.10%, respectively, with an FPR of 5%. The average execution time per vehicle image was 2.4530miliseconds.
The LBP_MHOG_SVM method yielded an excellent improvement compared with existing proposed methods, a 4% improvement for the FPR, and 1.50% for detection accuracy. The detection rate is 99.62%, with an FPR of 1.675%. The average of the processing time per vehicle image was 2.2187miliseconds.
Finally, the accuracy and detection rates are 97.92% and 99.71, respectively, with the FPR of 2.24% for the MRELBP_SURF_ELM method. The average of the execution time for the whole detection system per vehicle image was 2.108 milliseconds. This method was superior in the performance and execution time over the existing proposed methods in this research.
The findings suggest that the outcomes of this study can improve the performances of existing LPD systems. They can assist in law enforcement with an ITS system. Also, it can be effectively used to detect LPs in real-time applications under difficult conditions.
|ANZSRC Field of Research 2020||461206. Software architecture|
|460304. Computer vision|
|461204. Programming languages|
|400904. Electronic device and system performance evaluation, testing and simulation|
|460499. Cybersecurity and privacy not elsewhere classified|
|460905. Information systems development methodologies and practice|
|Byline Affiliations||School of Sciences|
0views this month
0downloads this month