Ferdous Sohel

Investigation of parts-based deep learning pipelines for identifying individual feral cats in camera trap imagery. Feral cats (Felis catus) pose a significant ecological threat in Australia, impacting native wildlife and contributing to species decline. Traditional monitoring methods like mark-recapture are limited in scope, leading to a focus on non-invasive techniques such as camera trapping and computer vision. These approaches promise efficient individual identification and provide insights into population dynamics but face challenges like variable image quality and difficulty in distinguishing individual animals. Previous studies in individual animal identification through imagery were mostly conducted in controlled or semi-controlled environments while focusing on specific animal features to discern individuals from one another. Imagery from automatic camera traps in uncontrolled environments often exhibit poor quality images and obscured or partial views of the animal, and their automatic capturing of animals in random poses and angle means specific features cannot be anticipated or chosen beforehand. We propose a parts-based approach for the identification of individual feral cats from automatic camera trap images, designed to accommodate partial or obscured views and utilise various distinguishable features of the cats. Utilising a subset dataset from a dataset of 15,881 camera trap images of cats from the Glenelg and Otway regions in Victoria, Australia, we propose a backbone ResNet for feature extraction and various model frameworks, including Mask R-CNN, Faster R-CNN, Inception, YOLO, and ResNet will be compared.