License plate recognition using capsule networks with an improved dynamic routing algorithm

다국어 초록 (Multilingual Abstract)

License Plate Recognition is a practical application based on computer vision. With the development of artificial intelligence and autonomous driving technology, this application plays a very big role in the system development. A traditional method is used by the convolutional neural network(CNN) for identification. But due to the intrinsic inability of max pooling layer, Convolutional Neural Network(CNN) fail to capture pose, view and orientation of the images. CNNs typically require a large amount of training data and cannot handle input process correctly. The Capsule Network is a recently introduced new machine learning architecture designed to overcome these disadvantages of CNN and contribute to a radical change in deep learning solutions. Because of robust to rotational and affine transformations of images, the capsule network is well suited to the processing of license plate image data sets.
In this paper, to improve the accuracy of license plate recognition, we adopt a novel method for this task using deep learning architecture called capsule networks. The capsule network consists of capsules, which are a set of neurons that use dynamic routing algorithm to represent object instantiation parameters, such as pose and orientation. First, in order to solve the problem that the license plate picture is difficult to collect, we designed a license plate image generation system to generate the data set. Second, we use the capsule network to implement the license plate recognition system, and design CNN as an experimental control group. Finally, we tested on the real license plate dataset. The result shows that the Capsule network’s accuracy rate on the test data set reaches 91.3%, and it is improved by 2.34% compared with the traditional convolutional neural network. In addition, we also designed a new capsule network improved dynamic routing algorithm for solving the problem that the capsule network training time is too long. And we also compare the performance of CNN and capsule network under the same training data set. Our results show that the training speed is increased by 31.9%.

목차 (Table of Contents)

• Chapter 1. Introduction 1
• Chapter 2. Related Work 4
• 2.1. Convolutional Neural Network 4
• 2.2. Capsule Network 8
• Chapter 3. Dynamic Routing Algorithm 12
• 3.1 Original Dynamic Routing Algorithm 12
• 3.2 Improved Dynamic Routing Algorithm 16
• Chapter 4. Experiment and Result 19
• 4.1. Data Set 20
• 4.2. CNN 23
• 4.3. Capsule Network 25
• 4.4. Result 26
• Chapter 5. Conclusion and Future Work 29
• References 30