An Improved Compression Layer Network Structure for VLPR

Abstract

License plate recognition is widely used in road conditions, parking, dredging traffic, etc. The license plate recognition first needs to find the specific location of the license plate and then needs to identify the content of the license plate as an image. During practical applications, main challenges of this recognition technology include the background color, size and specifications of the license plate, weather conditions, background interference, lighting, etc., which may also interfere with license plate recognition. Therefore, complex scenes and moving vehicles are two factors that need to be taken seriously. We propose an improved compression layer network structure for license plate recognition. The main way is to use a parallel structure of the compression layer network connection to alleviate errors caused by illumination, tilt and occlusion in license plate recognition. For reduce the amount of computation often pointed out in the CNN, we have tried to discard some of the parameters. Experiments have shown that this method has a lower error rate than CNN and capsule baseline, also has more advantageous in terms of time consumption. Introduction License plate recognition is an artificial intelligence technology that is fully utilized at home and abroad. It is widely used in automotive related fields such as traffic management, road tolls, ticket sales and charges [1]. The main techniques required for license plate recognition include segmentation, extraction and recognition of image processing techniques associated with license plates and the characters and symbols thereon. The researchers found many different ways to position license plates. For example, vector quantization based is the main scheme of traditional signal coding. In the license plate recognition system, VQ is used as a filter to identify the license plate area in the image as an image composed of certain lines and use the accumulation function to find various maximum values for the detection of the license plate area. Another kind of fuzzy logic algorithm firstly gives the rule and uses two methods of filtering and filtering to carry out vehicle license plate detection. All of these methods still suffer from the dilemma of light and background interference. Color is a feature that is often chosen, but the background interference cannot be eliminated. The edge features may be more suitable in this case. If the edge features can be improved and combined with the above features, it can give more the result of the identification of the value. The combination of edge and shape can be applied to extract the main area of the license plate. Experiments show that these features usually have good recognition performance [2]. Gradient and its variance are often used to identify, the main principle is that the brightness of the image in these areas is more obvious and rapid mutation. If the poor condition of illumination, occlusion, etc. interferes with the license plate, if the width and variance are added as supplementary features, the recognition performance can be further improved, and no more computing resources need to be introduced. However, this method is difficult to achieve the best results in complex road conditions or parking lots, because multiple changes in complex backgrounds in these scenes will also show rapid mutations in the image, resulting in a gradient and its variance. Increase in sex. However, this method is very suitable for real-time detection of license plates due to low computing resources and fast response. The traditional color license plate recognition system mainly adopts the following methods [3]: preprocessing, extraction of boundaries, segmentation, recognition, and post-processing. Among 2nd International Conference on Electrical and Electronic Engineering (EEE 2019) Copyright © 2019, the Authors. Published by Atlantis Press. This is an open access article under the CC BY-NC license (http://creativecommons.org/licenses/by-nc/4.0/). Advances in Engineering Research, volume 185