Use this URL to cite or link to this record in EThOS:
Title: Automated detection and shape based recognition of individual great white sharks
Author: Hughes, Benjamin
ISNI:       0000 0004 5994 4495
Awarding Body: University of Bristol
Current Institution: University of Bristol
Date of Award: 2016
Availability of Full Text:
Access from EThOS:
Systems that classify images by the individual animals they contain have widespread applicability in field-based ecology and conservation research, allowing individuals to be recognised repeatedly in a non-invasive way. The aim of this thesis is the design of such systems for animal species where individuals exhibit visually unique body morphologies, with the specific objective that individuals are recognised fully automatically. A two-stage approach is adopted to achieve this objective, as illustrated for the task of recognising individual great white sharks. First, a model is trained for automatic object part detection that combines a partitioning of ultrametric contour maps with shape descriptions and dense local features. This provides robust part detection but fine-grained segmentation accuracy is sacrificed in favour of computational efficiency. As such the approach is complemented by affinity matting for local edge refinement. The combination of part detection and affinity matting achieves robust, efficient and pixel accurate biometric contour detection. Second, a generative model combines evidence provided by densely sampled, multiscale local shape descriptions for biometric contour classification. The approach provides a discriminative representation of individuality while demonstrating robustness to sources of intra-individual variability introduced by partial occlusions and automatic shape detection errors. As an additional contribution, the distribution of individuality in dimensions of smoothing-filter scale, spatial location and descriptor complexity is quantified. Insights are provided to guide processes of image acquisition, shape representation, and efficient shape extraction. Finally, the generality of the contour representation is presented alongside a novel framework for discriminative cue combination in an application to individual humpback whale recognition. A detailed evaluation of the major system components is provided with results demonstrating fully automatic individual classification performance at accuracy and efficiency levels ready to assist human identification efforts.
Supervisor: Not available Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID:  DOI: Not available