Synthetic License Plate Generation: A Novel Approach For Effective License Plate Recognition In Malaysia

As deep learning algorithms devour large amounts of data, the quality and quantity of the dataset utilized to solve the problem of Automatic License Plate Recognition (ALPR) are critical to the success of any proposed solution. However, the variability in license plate designs, lighting conditions, vehicle position, and environmental factors leads to a real challenge due to the limitations of gathering a real dataset that can cover all these variations. This paper presents a novel approach for effective license plate recognition in Malaysia by generating a synthetic dataset for Malaysian license plates. The proposed approach leverages synthetic data generation techniques to create a diverse and comprehensive dataset of license plate images that capture the variability in Malaysian license plate designs and fonts. The synthetic dataset in machine learning is the process of generating artificial, unreal data that simulates real-world data. A fully automatic annotation method is also introduced to label the whole dataset without the need for the tedious manual labeling approach. By exploiting a novel text generation technique and systematically generating the character sequence, we generate realistic synthetic license plate images that closely resemble the characteristics of real license plates in Malaysia. The synthetic license plate images are automatically and accurately annotated using proposed novel algorithms. The two models are trained on a real dataset only and a combined dataset (real and synthetic). Experimental results demonstrate that the synthetic license plate dataset generation approach improves recognition accuracy compared to using only real license plate images with a whopping 12.1% performance increment.