DEEP LEARNING FRAMEWORK FOR CHARACTER RECOGNITION IN LOW QUALITY LICENSE PLATE IMAGES

Commercially available Automatic License Plate Recognition (ALPR) systems have limited ability to provide character recognition on low-quality license plate images [20]. Improving this ability would be beneficial for tasks currently requiring human involvement to read low-quality license plate characters. Recent advances in Deep Learning networks have shown that, for object detection tasks, Deep Learning networks can achieve levels of performance equal to or better than those of a human [2,6]. The aim of this thesis is to introduce a foundational Deep Learning framework for character recognition performance analysis. The analysis is carried out on license plate images that have undergone various types and levels of image quality reduction. This thesis leverages the TensorFlow Object Detection API to enable rapid development and testing of different Machine Learning networks and configurations. The framework allows for the creation of synthetically generated datasets on which image augmentation techniques can be applied. The various image augmentation techniques expand the dataset, and enable the network to be robust to image quality reductions. Networks were trained on the Maryland Advanced Computing Center’s GPU system. Per-character metrics of precision and recall are framework outputs used to evaluate trained networks. Network performance was evaluated using the framework for several Machine Learning models. The Faster R-CNN ResNet 50 network was found to have the best performance for character recognition on synthetically generated license plate images. On an ideal dataset, with no image degradation applied, the lower threshold of image size, on which the Faster R-CNN ResNet 50 network can reliably perform character recognition, was found to be 32 x 16 pixels. Finally, the network was trained and tested on image datasets with various data augmentation techniques applied. The data augmentation techniques evaluated in this thesis are: JPEG Compression, motion blur, affine transforms, and Gaussian noise. The results showed that, when trained on augmented synthetic data, the network was robust to quality reduction from most of the applied augmentation techniques