B1943
Title: Top-down classifiers with hierarchy based on taxonomic resolution used by human experts
Authors: Johanna Arje - University of Jyvaskyla (Finland)
Jenni Raitoharju - Finnish Environment Institute (Finland)
Alexandros Iosifidis - Aarhus University (Denmark)
Ville Tirronen - University of Jyvaskyla (Finland)
Kristian Meissner - Finnish Environment Institure (Finland)
Moncef Gabbouj - Tampere university of technology (Finland)
Serkan Kiranyaz - Qatar University (Qatar)
Salme Karkkainen - University of Jyvaskyla (Finland) [presenting]
Abstract: Our problem arises from bioassessments based on taxa recognition manually performed by experts. The practical target was to speed up taxa recognition by the methods of machine learning. For that purpose, we developed a software and technical prototype that allows for multiple images per specimen. Besides multiple images, the objects have hierarchical levels specifically created mimicking the classification strategy performed by human taxonomic experts. In that case, we constructed a top-down approach, that is, hierarchical classifiers built on support vector machines (SVM) and deep convolutional neural networks (CNN). The results were compared with typical flat classifiers and human experts using actual specimens. The lowest classification error 6.1\% was obtained by human experts, the second lowest 11.4\% and the third lowest 13.8\% by a flat CNN and a hierarchical CNN, respectively. Contrary to previous findings in the literature, we found that the flat classification approach commonly used in machine learning performs better than the hierarchical approach also called as a local per parent node approach. Moreover, we shared our unique dataset to serve as a public benchmark dataset in this field.
