Preliminary Study of Multi Convolution Neural Network-Based Model To Identify Pills Image Using Classification Rules

Abstract

Personal medicine is very important for those who have special health problems. Having several types of pills can make it hard for people to remember every pill especially aged citizen who easily forget his or her own medication. Another problem often encountered is the difficulty of recognizing the drug pills whose labels or the packaging are damaged and hard to read. This research, we developed a multi convolutional neural network (CNN) model to identify pills using classification rules. The idea of using multi CNN model is that almost all type of pills have three main identifiers, namely color, shape and imprint. Three CNNs model are developed to represent each identifier. The number of data collected is 24.000 images, which 95% of the data is used for training purpose and 5% is used for data test. The results of each CNN model is processed with some predefined rules to generate the classes of pills. From the results of different CNN architectures, number of epochs, optimizers and input size experiments, LeNet architecture with input size 64×64 pixels and Adadelta optimization shows the best accuracy up to 99.16%.