Reconhecimento de dígitos em imagens de medidores de consumo de gás natural utilizando técnicas de visão computacional

Abstract

This work proposes an approach that employs image processing techniques and pattern classification for the recognition of digits displayed in the counter of consumption of natural gas meters. These images are obtained in the field from actual operation conditions, unlike most of approaches in the literature that are based on images acquired in controlled environments. Initially, the counter of consumption is segmented by image processing techniques. The segmentation is performed based on the HSL color space of the image, unlike most of the works presented in the literature that uses images in grayscale. Then the digits are individually segmented and their features extracted in order to compose a knowledge base. This base serves as support to perform the classification and recognition of digits. Finally, a comparison is made between the performance of classifiers KNN, SVM and ELM in the recognition of segmented digits. The results demonstrate that the ELM neural networks have a performance superior to the other tested classifiers. Furthermore, the methodology showed to be promising in this scenario, reaching 95% success rate in recognition of digits. Presenting less than 5% fault in the process of segmentation the counter of consumption, considering a database with 903 images of gas meters. Given the scarcity of images compatible with the finality of this research, has as another objective of this work the availability of a database containing images of natural gas consumption meters. This base consists of images of gas meters with varying sizes, obtained from actual operating conditions. Also part of this base, resulting images of the individual segmented digits process, in order to meet researchers who wish to apply new methods of classification.