Optimizing Deep Learning for Efficient and Noise-Robust License Plate Detection and Recognition

Abstract

Accurate license plate recognition (LPR) remains a crucial task in various applications, from traffic monitoring to security systems. However, noisy environments with challenging factors like blurred images, low light, and complex backgrounds can significantly impede traditional LPR methods. This work proposes a deep learning based LPR system optimized for performance in noisy environments through hyperparameter tuning and bounding box refinement. We first preprocessed the noisy images by noise reduction which is crucial for robust LPR. We employed Convolutional Autoencoder (CAE) trained on noisy/clean image pairs to remove noise and enhance details. We utilized the InceptionResNetV2 architecture, pre-trained on ImageNet, for its strong feature extraction capabilities. We then added Region Proposal Network (RPN) head added to InceptionResNetV2 to predict candidate bounding boxes around potential license plates. We employed grid search to optimize key hyperparameters like learning rate, optimizer settings, and RPN anchor scales, ensuring optimal model performance for the specific noise patterns in the target dataset. Non-maximum suppression (NMS) eliminates redundant proposals, and a separate detection head classifies each remaining bounding box as license plate or background. Finally, bounding boxes are refined for improved accuracy. For confirmed license plates, a Bidirectional LSTM/CRNN network extracts and decodes character sequences within the refined bounding boxes.Compared to the recent methods, the proposed approach yielded the highest detection and recognition performance in noisy environments which can be best utilized for monitoring traffic, security systems in noisy environment. Our optimized LPR system demonstrates significantly improved accuracy and robustness compared to baseline methods, particularly in noisy environments.