EARL Research - ALPR for Bangladesh

ALPR: ResNet50 powered Bangla License Plate Detection and Optical Character Recognition

An algorithm to detect license plates with a novel architecture

Introducing a Novel Architecture for Bangla ALPR and OCR

We study the performance metrics of various optimization algorithms in order to determine the model that performs best on a novel dataset for Bangla License Plates. According to our study, the model performed best when ResNet50 was trained with the Root Mean Square Propagation Optimizer. We utilized Linear Support Vector Machines as the Classifier for our model.

IEEE Xplore Paper

ieeexplore.ieee.org/document/10543675

Abstract: This paper implements the MATLAB Image Processing Toolbox in detecting the license plate region using several user-defined functions in order to pre-process and process the image up until the point of extraction of characters. The extracted characters were then classified by utilizing ResNet50 from the Deep Learning Toolbox of MATLAB, custom training it on above a thousand images of Bangla and English characters and numbers alongside possible categories of noise extracted from the ROI after processing the image, which resulted in a datastore of 103 total categories. The output is converted to a string and saved in an Excel sheet to be accessed later on. In this ALPR model, the model will scan through the images of vehicles from a folder in a destination specified by the code to identify the license plates and characters and perform necessary actions on them. The aim of this paper is to properly implement the Image Processing Toolbox by MATLAB in order to identify the region of Interest and study the performance of the Linear SVM(Support Vector Machine) classifier with ResNet50 when it comes to Bangla OCR. The training and validation accuracy achieved by using the Root Mean Square Optimizer was 97.57%. The final accuracy and precision achieved while testing the model on 50% of the image dataset was 99.2%. Moreover, the ER (Error Rate) and FPR (False Positive Rate) were limited within 0.02%. The model scored 100% on F1 scores and Matthews Correlation Coefficient for every category of image classified.

Tools

Matlab Image Processing Toolbox

Image Deblurring, Noise Removal, and Image Sharpening functions.

Matlab Deep Learning Toolbox

ResNet50, linear SVM classifier, optimizers (adaptive momentum, root mean square propagation, stochastic gradient descent momentum, etc).

Metrics Studied

Error Rate

Error rate is the proportion of incorrect predictions made by a classifier compared to the total number of predictions. It gives a general idea of how often the model makes wrong predictions.

False Positive Rate

The False Positive Rate is the proportion of negative instances that were incorrectly classified as positive (i.e., false positives) out of all actual negatives. It measures the likelihood that a model incorrectly signals a positive detection when it shouldn’t.

F1 Score

The F1 score is the harmonic mean of precision and recall, providing a single metric that balances the trade-off between precision and recall. It is useful when you want to balance both false positives and false negatives.

Matthews Correlation Coefficient

The Matthews Correlation Coefficient is a balanced measure that accounts for true and false positives and negatives. It is often regarded as one of the best ways to evaluate binary classification models, especially when the classes are imbalanced.

Our story

A brief history

We started with the goal of automating license plate recognition technology in 2021 with the motivation of reducing overspeeding and illegal overtaking on the highways of Bangladesh. This work was a part of our final year thesis at Leading University, Department of Electrical and Electronic Engineering. Evaluation of metrics is important in any research project in order to understand the best combination of parameters for a model on a given dataset. In order to ensure the robustness of our model, we studied the performance metrics in the first 6 months of 2021 and decided to publish our findings in the 9th International Conference for Convergence in Technology held in India.

Team

Outing on the night of Ramadan

Hierarchy

Our team was led by A.N.Chowdhury and supervised by G.M.Chowdhury. S.P.Summit and M.F.A.Lasker contributed to all aspects of the work throughout the process. The research project was co-supervised by I.A.Chowdhury.