An Efficient Framework for Animal Breeds Classification Using Semi-Supervised Learning and Multi-Part Convolutional Neural Network (MP-CNN)

S Divya Meena,L Agilandeeswari

doi:10.1109/access.2019.2947717

Abstract

The automatic classification of animal images is an onerous task due to the challenging image conditions, especially when it comes to animal breeds. In this paper, we built a semi-supervised learning based Multi-part Convolutional Neural Network (MP-CNN) that classifies 35,992 animal images from ImageNet into 27 different classes of animals. The proposed model classifies the animals on both generic and fine-grained level. The animal breeds are accurately classified using Multi-part Convolutional Neural Network with a hybrid feature extraction framework of Fisher Vector based Stacked Autoencoder. Furthermore, with Semi-supervised learning based pseudo-labels, the model classifies new classes of unlabeled images too. Modified Hellinger Kernel classifier has been used to re-train the misclassified classes of animals and thereby improve the performance obtained from MP-CNN. The model has experimented with varied tasks to analyze its performance in each of the cases. The experimental results have proved that the coalesced approach of MP-CNN with pseudo-labels can accurately classify animal breeds and we have achieved an accuracy of 99.95% from the proposed model.

Highlights

Despite being the oldest computing technique, image classification remains an indispensable one
To overcome the above said shortcomings, we propose a novel Multi Part Convolutional Neural Network (MP convolutional Neural Network (CNN)) with both object localization and part selection models without any annotations
We have proposed a coalesced approach of Semi-supervised learning based Multi-part Convolutional Neural Network built on Tensorflow, which can classify animals on a fine-grained level

Summary

Introduction

Despite being the oldest computing technique, image classification remains an indispensable one. It has come a long way from using Fourier transforms to using neural networks. It remains a complicated computation because of the challenges in the images such as pose variations, occlusion, illumination, camouflage and more. A kind of machine learning lets the model perform classification directly from the training source like images, text, or sound. This requires the construction of a Deep Neural Network (DNN). One can use the concept of Transfer Learning to build very efficient neural networks

Objectives

Methods

Results