Implementation and Evaluation of a Simple Convolutional Neural Network for Object Classification in Visual Assistance Systems

Abstract

Image object recognition and classification has become more widespread in use and capable in functionality due to advancement in machine learning. One application of it is visual assistance systems for visually impaired persons, and there has been many existing proposed implementations and solutions about such systems. This article implements a simple convolutional neural network (CNN) machine learning model based on a simpler version of the architectures found in existing object detection models, which is trained using select data categories of the ImageNet dataset. The model is evaluated using sparse categorical accuracies, measuring the proportion of correct classifications, and comparing the number of predicted and expected classifications for each data category. The accuracy of the model did not perform well as expected because of significant overfitting behavior noticed through the validation loss. Even if early stopping to reduce overfitting is implemented, the overall accuracies still cannot be fully remediated. However, the difference between the number of expected and actual predictions is comparable.