Improving Optical Character Recognition in License Plate Images using Reinforcement Learning

Optical character recognition (OCR) is an essential component of computer vision systems, especially for tasks like license plate recognition. Problems arise when OCR is used on smaller pictures, which results in a significant loss in accuracy. Although modern technologies have attempted to overcome this obstacle, they face practical challenges, chief among which is the need for large datasets containing high-quality and low-quality pictures. This work presents a unique technique that uses reinforcement learning models to solve the challenge of improving OCR accuracy in lower-resolution pictures. Our methodology is motivated by the need to reduce reliance on large datasets using reinforcement learning. To stimulate this, we train our models using a small sample of license plates for training. This research paper compares the performance of different reinforcement learning models. Proximal Policy Optimization is identified as the best among the algorithms we tested and is on par with other research methods, even with smaller training data.