Orchids classification using spatial transformer network with adaptive scaling

• Main Authors: Watcharin Sarachai, Jakramate Bootkrajang, Jeerayut Chaijaruwanich, Samerkae Somhom
• Format: Book Series
• Published: 2020
• Subjects: Computer Science; Mathematics
• Online Access: https://www.scopus.com/inward/record.uri?partnerID=HzOxMe3b&scp=85076641558&origin=inward; http://cmuir.cmu.ac.th/jspui/handle/6653943832/67754
• Institution: Chiang Mai University

© 2019, Springer Nature Switzerland AG. The orchids families are large, diverse flowering plants in the tropical areas. It is a challenging task to classify orchid species from images. In this paper, we proposed an adaptive classification model of the orchid images by using a Deep Convolutional Neural Network (D-CNN). The first part of the model improved the quality of input feature maps using an adaptive Spatial Transformer Network (STN) module by performing a spatial transformation to warp an input image which was split into different locations and scales. We applied D-CNN to extract the image features from the previous step and warp into four branches. Then, we concatenated the feature channels and reduced the dimension by an estimation block. Finally, the feature maps would be forwarded to the prediction network layers to predict the orchid species. We verified the efficiency of the proposed method by conducting experiments on our data set of 52 classes of orchid flowers, containing 3,559 samples. Our results achieved an average of 93.32% classification accuracy, which is higher than the existing D-CNN models.