Detecção de landmarks e classificação de subespécies de abelhas através de asas com deep learning

Abstract

The aim of this project is to help bee hive maintenance in beekeeping to develop an automatic bee subspecies classifier. For this, a program was developed that uses the landmarks adapted from Nawrocka [1] for classification, but to perform the process automatically, it was necessary to implement an object detector capable of finding bee wings in an image and a detector of landmarks able to identify them in an image and then proceed to classification. The object detector was able to detect 98 % of the wings and the landmarks detector obtained was able to detect all landmarks in 91 % of cases, with a precision of 94 % resemblance to hand-marked landmarks. The classification, in turn, presented good results with the largest classes of datasets, with 92 % accuracy with the two largest classes and 87 % accuracy with the three largest.