Optical character recognition on electrical specification plates

Abstract

This paper looks into the feasibility of using neural networks to classify characters on electrical specification plates (ESP). This thesis is given by Verico AS (Verico). Verico performs large scale Asset Documentation, where photo documentation is one of their main tools. As such, they have large amounts of ESP imagery. Collecting data from the images is done manually, which is both time consuming and tedious work. This thesis seeks to further develop and utilize previous work done for Verico, such as background segmentation and vertical histogram analysis. The scope of the thesis is looking at the feasibility of using neural networks as a classifier for digits. MATLAB’s Neural Network Toolbox is used to train and classify data. The neural network is trained on 240318 images from the Street View House Numbers (SVHN) Dataset, and then tested on two different datasets. The first is on 26032 images from the SVHN test dataset, where the neural net achieved an overall accuracy of 84.1%. Through confidence thresholding 98% accuracy is reached at 52.8% coverage. The other dataset consists of 600 images gathered from several classes of ESP. The neural net achieves 94.1% overall accuracy, and with confidence thresholding 98% accuracy is reached at 85.3% coverage.