Transfer Learning applied to a Classification Task: a Case Study in the Footwear Industry

Abstract

Convolutional Neural Networks are a widely used method for image classification. They are part of the Deep Learning area, whose main advantage is the fact that they do not require an human support to extract features from the images. In the context of the footwear industry, they represent a useful computational resource, being applied for style classification problems, machine vision, among others. This article aims to evaluate the performance of transfer learning methods for the purpose of hierarchical classification of new items of footwear products. For this purpose, a dataset composed by 5,177 images of womens shoes was built. A pretrained architecture was selected to be refined, in order to produce a classification model. As a main result of this study, we confirm that the use of transfer learning speeds up deep neural nets training, allowing outstanding results through a VGG16 architecture. In terms of accuracy, the results achieved 99.97% and 98.42% for classifying respectively footwear categories and subcategories.