Automated tracking of insects

Details

Supervisors

Bonaventure, Olivier ; Detry, Renaud

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] : ingénieur civil en informatique, à finalité spécialisée

Abstract

Insect identification is a complex task that is mainly executed by professional entomologists that manually distinguish their different characteristics mainly by the use of an identification tree. Our work aims to facilitate this exercise by automating it with the use of neural networks. The goal of this thesis is to provide an automated insect detection and classification trap system that focuses on ten different Hymenoptera species such as Apis, Bombus, Anthophora and Vespula species. This master’s thesis studies the different solutions in the literature and compares their advantages and disadvantages with respect to the constraints of the project, namely the Faster R-CNN and Mask R-CNN architectures. We inspected similar works in the scope of insect identification to guide our conception choices. We decided to use the Faster R-CNN architecture, as it best matched our constraints, to train a 10 classes model, and used several image augmentation techniques that increased the model’s quality during the validation phase. Finally, we deployed our model on an embedded insect trap system to produce detections. Our model reached an AP score of 73.9% and a ROC AUC value of 71.9% when evaluated on generic data, proving that satisfactory results can be produced with this solution. The precision and recall values of the final configuration that aims to maximize the F1-score reached the values of 76.8% and 74.1% respectively. The inference on experimental data from the trap resulted in a mean precision and recall for each insect of 57.3% and 53.7% respectively, but that data was really limited as the data harvesting phase was restricted by time and weather constraints. We could however show that it is possible to build a heuristic based on our model’s output that could theoretically reach higher values of precision and recall using the hypothesis that the mean precision for each insect is higher than 50% and that consecutive images often correspond to the same insect.