A Simple Strategy for Choosing Network Structures in a Object Detection Project with Transfer Learning

Abstract

Compare different network configurations in the early stages of an object detection project can be an interesting approach to identify the one that can provide the best performance and, thus, optimize the investment of time and research efforts for the next steps. In this work we will explore the issue through the study of object recognition applied to a category of items, specifically fruits, where the proposed strategy will be to select a public image dataset of these items and to train some different structures of deep learning networks. We built different combinations of structures composed of pre-trained base networks, in which the upper layers were replaced by new structures, with an increasing degree of complexity. Then will evaluate the results of these pre-trained networks with 25 images of individual fruits obtained on the internet. After we compare the performance between the different structures of networks, it is intended to demonstrate if there is a relationship between the training performance of specific models with the complexity of its upper layers when we apply them to a practical evaluation.