Character Recognition Architecture for Low-Cost Glucometers: A Step Towards the Deployment of an Ehealth System in Colombian Rural Communities

This work presents a digit recognition algorithm designed for low-cost glucometer displays, leveraging the pattern retrieval capabilities of Hopfield and Convolutional Neural Networks. The system processes input images through grayscale conversion, adaptive binarization, and morphological filtering to enhance digit visibility. Each digit is segmented, resized, and transformed into a binary vector suitable for associative memory recognition. A pre-trained Hopfield network retrieves the closest stored pattern by converging to a stable state, effectively identifying each individual digit; this output is then sourced to a Convolution Neural Network that let us assemble meaningful data matching for glucometric information, date and hour of measurement. The system was modeled by Unified Modeling Language diagrams with a pseudocode and successfully tested on Glucoquick glucometers, which are popular in rural areas of Colombia, under conditions of daylight and artificial lighting.