Chinese vehicle license plate recognition using kernel‐based extreme learning machine with deep convolutional features

Abstract

License plate recognition (LPR) is an important component of intelligent transportation systems. Compared with letters and numbers, Chinese characters contain more information, making automatic recognition more difficult. Accurate Chinese LPR (CLPR) is determined by three factors: training dataset, feature extractor, and classifier. Most license plates with benchmark dataset contain only letters and numbers; thus, the authors build a large dataset for CLPR. Convolutional neural networks (CNNs) can be used to extract inherent image features, on all levels of abstraction. CNNs can be used for classification if they have a sufficient number of fully connected layers. This implies that CNNs must be trained using gradient descent-based methods, which often yields sub-optimal results. Extreme learning machines (ELMs) demonstrate impressive performance on classification, with good generalisation. Therefore, the authors propose a novel deep architecture for CLPR which combines a CNN and an ELM. Firstly, a CNN without fully connected layers, working as a feature extractor, learns deep features associated with characters in written Chinese. Then, a kernel-based ELM (KELM) classifier, which accepts CNN features as input, is utilised for classification. Compared with CNNs that use Softmax, support vector machines and ELMs, the CNN that uses KELM yields competitive results in a shorter training time.