Cross-Domain Plant Classification with Triplet Networks

Automated plant recognition using deep learning remains challenging, specifically for species that lack training images such as those in tropical regions. Herbaria (dried plant specimens) images have been adopted to mitigate this shortfall since they are more accessible, however, it is hindered by the plants' appearance changes during the drying process. This research addresses the issue by exploring how deep convolutional neural network can be constructed to integrate images of plants in the field with herbaria for model training. The devised convolutional neural network has shown to better encapsulate herbarium-field features than conventional neural networks, making it an important reference.