Neural networks with lstm and gru in application to the task of multiclass classification of text posts of social network users

Abstract

A lot of research has already been done in this area, and various models have been proposed that solve the problem of license plate recognition, there are industrial designs used on cameras for recording traffic violations. However, such models are usually developed for client-server architectures, since model architectures that solve this problem well often have hundreds of millions of parameters [5] and are designed to be trained and applied on machines with large memory and productive video cards. The subject of the study of this article is the automobile state license plates in the video sequence of poor quality, recognizable under conditions of limited computing resources. To recognize license plates, it is proposed to use the following methods: simple convolutional neural networks to determine the angles of a license plate and its subsequent rotation; a pretrained YOLOv5s model for searching license plates in a frame and a pretrained SCR-Net network for recognizing license plates on plates. The article proposes an algorithm and its implementation for automatic license plate recognition under conditions of poor video quality and limited computing resources. The theoretical significance of the result lies in the development of a new algorithm that takes into account these limitations, which contributes to the development of a series of studies in the field of developing methods and algorithms that operate under conditions of significant limitations on computing power and video quality. The practical significance of the result lies in the application of the result to solving applied problems, such as fixing incorrect parking, searching for stolen cars, and other offenses.