Gradual Labeling of the Training Set to Improve the Efficiency of Image Detection by a Neural Network on the Example of License Plate Recognition

Currently, the automation of various tasks is very important in modern society. One of these tasks is the recognition of the car license plate in the video stream. Our project is aimed at this problem. This task is not new and there are many open source solutions available. All solutions can be two types of implementation: with using neural networks and without using it. Our implementation is a neural network for car license plate recognition on an image. We use the YOLOv3 detector for recognition. Training of neural networks is often complicated due to the small number of training samples. When you train a neural network using an insufficiently amount of data, it is difficult to achieve the required effectiveness. This problem can be solved in two ways: by expanding the training dataset, or by selecting the optimal training parameters and transforming the existing dataset. This article describes the solution to this problem by selecting the optimal training parameters and transforming the training dataset. The gradual labeling of the training set is discussed. The tuning of the anchor boxes parameter is discussed in detail. The presentation of the test results provides a visualization of the work performed. Research results show an increase in the efficiency of the neural network detector by 30.5% from the initial value. Submitted method helps to save a lot of time, because automate expanding the training dataset.