Dr. Watcharin Sarachai
Sign in
complete

Orchids Classification Using Spatial Transformer Network with Adaptive Scaling

A deep learning approach for classifying orchid species from images, combining an adaptive spatial transformer network with a convolutional neural network to improve recognition accuracy.

Orchids Classification Using Spatial Transformer Network with Adaptive Scaling

Overview

Orchids form one of the largest and most diverse plant families, and many species share very similar petal shapes and colors, making automated identification genuinely difficult. This project proposes an adaptive deep convolutional neural network built around a spatial transformer module. Instead of feeding the raw image straight into a classifier, the model first learns to warp and crop the input at four different scales and locations, letting it focus on the most distinctive parts of each flower. A shared CNN then extracts features from all four warped views, which are concatenated and compressed by a dimension-reduction block before being passed to the final prediction layer. Evaluated on a custom set of 52 orchid species with 3,559 images, the approach reached about 93% classification accuracy, outperforming standard CNN baselines that lack this adaptive attention step.

Authors

Watcharin Sarachai, Jakramate Bootkrajang, Jeerayut Chaijaruwanich, Samerkae Somhom

Publication

Lecture Notes in Computer Science, vol. 11871 — Intelligent Data Engineering and Automated Learning (IDEAL 2019). Published by Springer, Cham, 2019.

Zygopetalum orchid flower with distinctive petal and lip patterning
Zygopetalum orchid flower with distinctive petal and lip patterning

Orchid species like this Zygopetalum hybrid show the fine-grained petal and lip patterns that make automated classification challenging — the kind of detail the adaptive spatial transformer in this work learns to zoom into. Photo by Petar Milošević, CC BY-SA 4.0.