A Key Point-Based License Plate Detection with Pyramid Network Structure

Abstract

In this paper, a key point detection method of license plate based on convolution network is proposed. Traditional license plate detection methods use features like shape, texture and color to locate a license plate with defects such as pertinence, high time complexity, window redundancy and poor robustness. The license plate detection methods based on deep learning have been greatly improved in accuracy and real-time performance, but the detection results of license plates with large rotation angle, small size, less illumination and occlusion are poor. In our method, the rotation angle of the license plate is obtained by detecting four corners of the license plate, and the perspective transformation is used for correction. In order to improve the location accuracy of license plate object, this paper proposes a pyramid network structure to extract high-level and low-level semantic features. Experiments show that the proposed model can not only detect the license plate in general scenes, but also has good detection effect for license plate with large rotation angle.