Plataforma de evaluación de algoritmos de reconocimiento de caracteres numéricos en imágenes digitales.

Abstract

The research aimed to evaluate the techniques of digit recognition in images: K-Nearest Neighbor (KNN), TESSERACT and Artificial Neural Networks (RNA); Which allowed to determine the technique with greater level of precision that can be used in different applications of artificial vision. The research hypothesis "The proper selection of numerical carácter recognition technique in digital images will allow to determine the technique with the highest degree of accuracy and to minimize the consumption of resources in SBC computers." The nonparametric Kruskal-Wallis

test was used in the comparison of the degree of precision of the techniques with a confidence level of 95%. The indicators that were used are: degree of precision, time spent for recognition, amount of RAM and level of CPU usage. A sample was used of 30 photographs taken at electric power meters of Las Acacias in Riobamba city. According to the statistical tests results, was determined that the most accurate recognition technique was KNN reaching an average of 49.33% versus to 29.833% and 22.00% of TESSERACT and RNA respectively. The average time spent by KNN for the recognition was 1.22 seconds, on average 15.7 megabytes of RAM and 11.64% of CPU usage were used. OpenCV, Python and C ++ were used under the Linux Raspbian distribution to evaluate each selected techniques. It is recommended to deepen the study of the KNN technique focused on different applications of artificial visión in the industry, as well as to evaluate new recognition techniques of Pattern such as SVM, Deep Learning and Non-Supervised Leanring, which have become active research fields in many universities around the world.