Dr. Watcharin Sarachai
Sign in
complete

Orchid Classification Using Homogeneous Ensemble of Small Deep Convolutional Neural Network

A hybrid deep learning approach combining a global feature network with a local, spatial-transformer-based feature network, fused through an ensemble of small CNNs, to better distinguish visually similar orchid species from images.

Orchid Classification Using Homogeneous Ensemble of Small Deep Convolutional Neural Network

Overview

Orchid flowers from different species can look strikingly similar, making automated species identification a genuinely difficult computer vision problem. This project proposes a hybrid architecture that tackles the problem from two angles at once: a global prediction network (GPN) that reads overall flower shape and color, and a local prediction network (LPN) that zooms into specific plant organs using a spatial transformer network. An ensemble neural network (ENN) then fuses the predictions from both networks, using several small convolutional neural networks together rather than one large model, to produce a more accurate final classification.

Authors

Watcharin Sarachai, Jakramate Bootkrajang, Jeerayut Chaijaruwanich, Samerkae Somhom

Publication

Machine Vision and Applications, Volume 33, Issue 1, 2022. Published by Springer Berlin Heidelberg.

Diagram of a convolutional neural network architecture
Diagram of a convolutional neural network architecture

A convolutional neural network architecture, showing feature learning through convolutional and pooling layers followed by a classification stage — the same general building block combined into an ensemble in this research.