Master thesis : Deep-Butterflies : Automatic Landmark Detection

Abstract

In the biomedical research field, and more precisely the field of morphometric studies, the detection of anatomical landmarks is a crucial step in order to quantify the shape and size of an object under study. The annotations of theses landmarks is quite laborious and often requires dedicated human expertise. Thus the use of automatic landmark detection techniques using artificial intelligence began to gain importance.

This thesis explores different kinds of approach in order to tackle the problem on butterflies from the Morpho genus. Each butterfly possesses two types of landmarks: true landmarks and semi-landmarks, and has two views: ventral and dorsal, resulting in a total of 4 distinct datasets. Moreover, the Morpho genus contain about 30 species which induces a great variety between the images.

Several approaches have been experimented, namely computer vision, machine learning and deep learning. The results shows that the deep learning approach outperforms the others in most cases. Nevertheless the machine learning approach has proven its performance on smaller part of the data, \textit{i.e.} datasets composed of one specie only. Unfortunately, the computer vision approach did not lead to any convincing results.

Finally, this work presents the Cytomine application that has been built along with the most consistent model from the experimentations. This application provides a user-friendly interface for both training and predicting with a new model.