Convolution Neural Networks: A Comparative Study for Image Classification
Narayana Darapaneni, Balaji Krishnamurthy, Anwesh Reddy Paduri
Abstract
Wide range of convolution neural network architectures are available for image classification, segmentation and object detection. Most of the architecture focus on accuracy as primary factor for implementation. However, when it comes to real time application deployment, there are other primary factors like memory and performance which is equally important. Also, each CNN architecture showcases its advantages and limitations but comparison over their peers are not equally considered. The goal of this paper is to provide a comparative study of various CNN architecture for image classification and serve as a guide for selection based on applications requirement and hardware capabilities. In this paper, we discuss about 18 different CNN state of art architectures that are widely used. In order to access model suitability for a given problem, CIFAR-10 image dataset is trained on different architectures with a specified set of hyper-parameters to measure the accuracy, performance and memory consumption. The experiment findings are presented to suggest suitable CNN architecture based on application/hardware attributes.