LICENSE PLATE NORMALIZATION BASED ON 3D NEURAL NETWORK

Saitov I. A., Filchenkov A. A.

In license plate recognition systems, an important quality aspect is the position of the cameras, which are usually located on the side or on top of the vehicles. Thus, the license plate image exhibits various distortions that affect the optical character recognition. The normalization of rotation angle distortion based on computer vision techniques is an additional step in number plate recognition algorithms. We propose an approach that uses a 3D vehicle detection model to find the object's deflection angle for subsequent perspective transformation of the number plate. A dataset of license plate images belonging to previously predominantly CIS countries is used as data. Character recognition experiments using the proposed normalization show its effectiveness, achieving an accuracy value of 0.939 on the considered data.

Keywords:neural network, license plate recognition, 3D detection, optical character recognition, orientation normalization