Recognizing Low Quality Vehicle License Plates using Image based Sequence Recognition

In present research work, we focus on design and development of a deep neural network to recognize low quality real world vehicle license plates. In the most of the previous approaches, different components of neural network are either trained or fine tuned separately. We make use of a unified architecture to integrate feature extraction, feature encoding-decoding for novel end-to-end neural network training. Feature encoding-decoding using recurrent neural network processes license plates with arbitrary lengths and provides an efficient compact model. We evaluate performance of our different models on synthetic and real world datasets. We achieve more than 99.9% accuracy for multi line, multi style, and variable length synthetic image dataset with more than 1.9% improvement in accuracy than previous works. We realize more than 98.5% accuracy for low resolution and low quality real world image dataset, which is similar to previous work.