Logo and Brand Recognition from Imbalanced Dataset Using MiniGoogLeNet and MiniVGGNet Models

Abstract

Deep learning model tends to promote models with deep structure. Despite its high accuracy, the model was not practical when high computing power was not available. Thus, deep model with not-so-deep structure or less number of model parameters is needed for low capacity computer. Logo and brand recognition task is an important and challenging problem in computer vision with wide potential applications. The inherent challenge to address this task is not only due to the presence of logo in various direction and clutters as well as imbalanced dataset but also because of high computing workload when deep learning models were adopted. This paper presents empirical results of logo recognition method using MiniVGGNet and MiniGoogleNet models combined with augmentation technique to increase variation and number of samples. The results show that the proposed model combined with augmentation technique increased accuracy of model accuracies and fasten training convergence of both models.