License Plate Detection And Recognition In Complex Scenes

As an important part of intelligent transportation system, license plate detection and identification technology plays a vital role in road traffic flow monitoring, traffic accident scene investigation, automatically record traffic violations, intelligent vehicle target recognition and other aspects of the public security department. However, due to the complexity of scenes and increased suitability requirements for more light, multi-angle and variable scale, license plate detection and recognition technology is also facing many problems to be solved. This thesis focuses on how to achieve steady license plate detection and license plate character recognition in multi- light, multi-angle, multi-scale complex scenes. The key elements are as follow:1. To solve the problem that character regions are difficult to be de tected in complex environment where light and angles change and the size of character regions vary, this thesis improves the stable license plate character region location algorithm. Combining the prior knowledge of license plate characters, this algorithm extract the extreme stable character regions of images by improving MSER core algorithm. Meanwhile, compared to the stable regions extracted by MSER, this algorithm would filter out a large part of pseudo-character region, which would reduce the workload of tags conduction in later processes.2. On how to accurately construct license plate character structure of tags conduction in case of multiple extreme stable character regions, this thesis proposes the largest movable structure chain algorithm. This algorithm conducts a crude extraction of extreme stable character regions through node classification, after which it achieves tags largest group extraction through semi-conductive tag label. Finally it applies the maximum posterior probability conditions analysis on tags largest group through full tag label conduction so as to correct tags and obtain the license plate location.3. To solve the problem of low recognition rate of low-resolution images, we adopt the methodology of convolution neural network arc hitecture character recognition. This method first preprocesses and normalizes character images. With self- learning neural network training features, it conducts character recognition with character grayscale image as input. Finally the character images with recognition result of similar characters are subject to second character recognition, which distinguishes between similar characters to gain the final result of character recognition.4. Design and implement multi-source input simulation presentation so ftware, which has single image, multiple images and video input ports, and integrated license plate detection and character recognition algorithms, to display visualization demos.