Image similarity usage in order to find similar items in E-commerce dataset

Abstract

This thesis looks into the e-commerce situation among small businesses and examines the possibility of the application of neural networks to business-to-consumer interactions. The goal of the study is to test the ability of NN to work with limited data and hardware. Pre-trained ResNet-50 model was used as a base, and fine-tuned with the 4 class dataset, consisting of 2906 images of clothes categorised into four classes. The experiments were conducted on a budget-friendly hardware setup. Both the dataset and hardware reflect the limited resources of the small and micro businesses The final model was evaluated, reaching an accuracy of 70-80%, These results suggest the possibility of NN usage in limited data and computational resource conditions and provide the base for further developments in implementing NN among small businesses.