This paper studies convolutional neural networks (CNN) to learn unsupervisedfeature representations for 44 different plant species, collected at the RoyalBotanic Gardens, Kew, England. To gain intuition on the chosen features fromthe CNN model (opposed to a 'black box' solution), a visualisation techniquebased on the deconvolutional networks (DN) is utilized. It is found thatvenations of different order have been chosen to uniquely represent each of theplant species. Experimental results using these CNN features with differentclassifiers show consistency and superiority compared to the state-of-the artsolutions which rely on hand-crafted features.
展开▼
