Research Area:  Machine Learning
Deep learning has obtained many successes in different computer vision tasks such as classification, detection, and segmentation of objects or faces. Many of these successes can be ascribed to training deep convolutional neural network architectures on a dataset containing many images. Limited research has explored deep learning methods for performing recognition or detection of animals using a limited number of images. This thesis examines the use of different deep learning techniques and conventional computer vision methods for performing animal recognition or detection with relatively small training datasets and has the following objectives: 1) Analyse the performance of deep learning systems compared to classical approaches when there exists a limited number of images of animals; 2) Develop an algorithm for effectively dealing with rotation variation naturally present in aerial images; 3) Construct a computer vision system that is more robust to illumination variation; 4) Analyse how important the use of different color spaces is in deep learning; 5) Compare different deep convolutional neural-network algorithms for detecting and recognizing individual instances (identities) in a group of animals, for example, badgers. For most of the experiments, effectively reduced neural network recognition systems are used, which are derived from existing architectures. These reduced systems are compared to standard architectures and classical computer vision methods. We also propose a color transformation algorithm, a novel rotation-matrix data-augmentation algorithm and a hybrid variant of such a method, that factors color constancy with the aim to enhance images and construct a system that is more robust to different kinds of visual appearances. The results show that our proposed algorithms aid deep learning systems to become more accurate in classifying animals for a large number of different animal datasets. Furthermore, the developed systems yield performances that significantly surpass classical computer vision techniques, even with limited amounts of available images for training.
Name of the Researcher:  Emmanuel Okafor
Name of the Supervisor(s):  Schomaker, Lambert
Year of Completion:  2019
University:  University of Groningen
Thesis Link:   Home Page Url