Automated method for detection and recognition of seven-segment digits from electric meters using machine learning
Loading...
Date
2022
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
UMT Lahore
Abstract
Technology is growing rapidly and everything should be automated to save time, cost, and effort. This research aims to accurately detect and recognize seven-segment digits from captured images of digital electric meters in Pakistan. Detecting and recognizing seven-segment digits from collected data samples may be helpful to improve the infrastructures of our metering systems. In this research, Dataset is collected from the power development authority officials, containing 23,000 sample images of electric meters. These samples are taken under multiple daylight conditions. Some samples are blurred, having shadows, poor contrast, and tilted which makes it more challenging to accurately recognize digits. First, we used Edge detection to find our region of interest (ROI), an LCD of a meter. After extracting ROI, Computer vision (CV) techniques such as gray scaling, Gaussian blur, adaptive thresholding, and some morphological operations are applied. After all of this preprocessing, digits are then segmented based on the features and pixel values. Finally, these segmented images (DIGITS) are then given to our pre-trained Convolutional Neural Network (CNN) model. This CNN model is trained on 24,475-digit samples which are capable of giving 92% accuracy. Data augmentation is also applied while training the model. Multiple machine learning algorithms are used like k-nearest neighbors (KNN), decision tree (DT), random forest (RF), and Support Vector Machine (SVM) in this research. Results show that using a Support Vector Machine (SVM) gives better results than using other models.