Detection and segmentation of fauna in seafloor imagery for biomass estimation

Abstract

Machine learning based image processing is sensitive to variation caused by hardware and observation conditions, making the use of machine learning with marine imagery particularly difficult when transferring knowledge between datasets. There is also a considerable gap between the outputs of machine learning systems and useful biological information for marine conservation purposes. This thesis investigates the effects of physics-based image normalisation and augmentation methods on the transferability of an object detection and segmentation system between two distinct datasets taken at different altitudes from the seafloor with different camera and lighting systems. Scale, colour, and lens distortion correction methods are investigated, along with augmentation methods including linear contrast, motion blur, and noise, and more advanced distorting methods such as elastic distortions and piece-wise affine transformations. A set of experiments for each combination of independent variables has been carried out, finding a clear improvement when using scale correction. When applying to low altitude datasets only there is an increase in average performance from 62.2% to 68.6%, and when transferring knowledge from high to low altitude datasets, there is an increase in performance from an average of 26.5% to 44.1% when using scale normalisation. Colour normalisation also had a large impact, when applied to low altitude data showing an increase in performance from 56.6% to 74.1%, and when transferring from high altitude to low altitude datasets showing an increase in performance from 32.7% to 38.0%. The impacts of lens distortion correction and various augmentation methods were found to be less significant. This thesis goes on to demonstrate the use of segmentation results for biomass estimation through a simple polynomial relationship between segment size and length of an individual, and previously well-established Length Weight Relationships (LWRs). The resulting method is fully scalable to larger datasets with no additional human effort required, a vast improvement on the current labour-intensive biomass estimation methods used.

More information

Identifiers

ORCID for Jennifer Walker: orcid.org/0000-0002-1449-9012

Catalogue record

Export record