Recognizing trees at a distance with discriminative deep feature learning

Abstract

We investigate discriminative features that are able to improve classification accuracy on visually similar classes. To this end, we build a deep feature learning network, which learns features with discriminative constraint in each single layer module, and learns multiple levels of features for hierarchical image representation. Specifically, the network encodes the discriminative information by automatically selecting the informative features, and forcing them to be closer to the features extracted from the same class than the features from different classes. We also collect a new fine-grained dataset containing 51 common tree species in Singapore. All the images are taken at a distance with large intra class variance, which makes the tree species hard to be distinguished. Our experimental results show that we are able to achieve 78.03% in accuracy on this challenging dataset, which is 8.48% higher than general hand-designed feature.