CNN filter sizes, effects, limitations, and challenges: An exploratory study.

Abstract

This study explores the impacts of filter sizes on convolutional neural networks (CNNs) models, moving away from the common belief that small filters (3x3) give better results. The goal is to highlight the potential of large filters and encourage researchers to investigate their capabilities. The usage of large filters will increase the computational power which leads common researchers to reduce the filter size to reserve this power; however, other researchers address the potential of large filters to enhance the performance of CNN models. Currently, there are few pure CNN models that achieve optimal performance with large filters showing how far the large filter sizes topic is not addressed well by the community. As the availability of computer power and image sizes increase, traditional obstacles that hinder researchers from using large filter sizes will gradually diminish. This paper guides researchers by analysing and exploring the limitations, challenges, and impacts of CNN filter sizes on different CNN architectures. This will help utilize large filters' distinctive opportunities and potential. To our knowledge, we find four opportunities from utilizing large filters. A comprehensive comparison of researches on different CNN architectures shows a bias for small filters (3x3) and the possible potential of large filters.