Методика предобработки данных машинного обучения для решения задач компьютерного зрения

Abstract

In the field of machine learning, there is no single methodology for data preprocessing, since all stages of this process are unique for a specific task. However, a specific data type is used in each direction. The research hypothesis assumes that it is possible to clearly structure the sequences and phases of data preparation for text recognition tasks. The article discusses the basic principles of data preprocessing and the allocation of successive stages as a specific technique for the task of recognizing ABC characters. ETL set images were selected as the source data. Preprocessing included the stages of working with images, at each of which changes were made to the source data. The first step was cropping, which allowed to get rid of unnecessary information in the image. Next, the approach of converting the image to the original aspect ratio was considered and the method of converting from shades of gray to black and white format was determined. At the next stage, the character lines were artificially expanded for better recognition of printed alphabets. At the last stage of data preprocessing, augmentation was performed, which made it possible to better recognize ABC characters regardless of their position in space. As a result, the general structure of the data preprocessing methodology for text recognition tasks was built.