The Research of Chinese License Plates Recognition Based on CNN and Length_Feature
Although the license plate recognition system has been widely used, the location and recognition rate is still affected by the clarity and illumination conditions. A license plate locating (LPL) method and a license plate characters recognition (LPCR) met
- PDF / 712,027 Bytes
- 9 Pages / 439.37 x 666.142 pts Page_size
- 15 Downloads / 155 Views
Harbin Institute of Technology Shenzhen Graduate School, Shenzhen, China [email protected] 2 College of Information Science and Engineering, Fujian University of Technology, Fuzhou, China
Abstract. Although the license plate recognition system has been widely used, the location and recognition rate is still affected by the clarity and illumination conditions. A license plate locating (LPL) method and a license plate characters recognition (LPCR) method, respectively, based on convolution neural network (CNN) and Length Feature (LF), are proposed in this paper. Firstly, this paper changes the activation function of CNN, and extracts local feature to train the network. Through this change, the network convergence has sped up, the location accuracy has improved, and wrong location and long time consuming, which caused by some complicated factors such as light conditions, fuzzy image, tilt, complex background and so on, have been resolved. Secondly, the LF, which is proposed in this paper, is easier to understand and has less calculation and higher speed than transform domain features, and also has higher accuracy to recognize fuzzy and sloping characters than traditional geometric features. Keywords: License plate location · Convolution neural network License plate characters recognition · Length Feature
1
·
Introduction
As an important part of public transport management, license plates recognition (LPR) reduces the complexity of traffic management, and improves the traffic capacity and management efficiency. There are four main technologies, including LPL, characters correction, cutting and LPCR, in a practical LPR process. Up to now, many valid methods have been proposed in these four main technologies. In LPL, there are some general methods based on different features and algorithms such as color, texture, transform domain, neural network. The color features in RGB could describe color information of license plates well, and it is enough to locate license plates in good bright images [9,16]. The result after discrete wavelet transform (DWT) is an effectual information to describe images [22]. A sliding concentric window was applied to scan images to find license plates in Giannoukos and Anagnostopoulos [7]. Some morphological methods, c Springer International Publishing Switzerland 2016 H. Fujita et al. (Eds.): IEA/AIE 2016, LNAI 9799, pp. 389–397, 2016. DOI: 10.1007/978-3-319-42007-3 33
390
S. He et al.
such as sobel edge detection, skeleton, region growing, could combine to find license plates well [11]. In license plate character segmentation, there are the projected image analysis and connected area segmentation, etc. The simplest one is segmenting characters depending on width of characters and distance between characters, but the simplest is the most inflexible, which requires accurate plate region [18]. Analyzing projected image of license plate should find the right border of every characters [15]. A character segmentation technique based on visiting neighbor pixel algorithm, which uses the connectedness of
Data Loading...
