Description
A Master of Science thesis in Computer Engineering by Ali Reza Sajun entitled, “Exploring Semi-Supervised Learning Algorithms for Camera Trap Images”, submitted in August 2022. Thesis advisor is Dr. Imran Zualkernan. Soft copy is available (Thesis, Completion Certificate, Approval Signatures, and AUS Archives Consent Form).
Abstract
Animal Extinction and biodiversity loss is a major challenge faced by ecologists in today’s world due to increasing populations and industrialization. In their fight to protect endangered species, ecologists use remote camera traps that allow them to monitor animals in remote locations. A drawback of current camera traps is that captured images need to be manually labeled through an error-prone, tedious and expensive process. Images from camera traps can be automatically labelled using deep learning techniques. However, these images are expensive to label, and the data is highly unbalanced. Semi-supervised learning can be utilized to address the lowlabelling issue, but the problem of handling unbalanced data remains unexplored. This thesis explored how state-of-the-art semi-supervised learning algorithm like FixMatch based on consistency regularization and pseudo-labeling and its variants including Auxiliary Balanced Classifier (ABC), Distribution Aligning Refinery of Pseudo-label (DARP) and Bi-sampling Strategy (BiS) performed on such highly unbalanced data. Imbalance ratios of 1, 50, 100 and 150 across labeled proportions of 10%, 40%, 60%, and 80% were considered. Additional experiments were also conducted on the benchmarking datasets of CIFAR10, CIFAR100 and SVHN. The primary results are that the resampling strategy worked best when applied to FixMatch for low levels of imbalance across the data sets. For example, an F1-score of 68% for CIFAR10 at 40% labeled data was achieved as opposed to an F1-score of 60% with the plain FixMatch. However, addition of auxiliary loss or pseudo-label balancing showed negligible improvements. The results also suggest that a major factor mediating the performance was complexity of the features. In general, the semi-supervised techniques performed well for the camera trap data with F1-scores of up to 58%. However, the overall performance on the camera trap data was affected by classes belonging to very small animals which occupied a small subset of the image frame and, therefore tended to be misclassified.