Automated interpretation of benthic stereo imagery

摘要

Autonomous benthic imaging, reduces human risk and increases the amount of collected data. However, manually interpreting these high volumes of data is onerous, time consuming and in many cases, infeasible. The objective of this thesis is to improve the scientific utility of the large image datasets.
\nFine-scale terrain complexity is typically quantified by rugosity and measured by divers using chains and tape measures. This thesis proposes a new technique for measuring terrain complexity from 3D stereo image reconstructions, which is non-contact and can be calculated at multiple scales over large spatial extents. Using robots, terrain complexity can be measured without endangering humans, beyond scuba depths. Results show that this approach is more robust, flexible and easily repeatable than traditional methods.
\nThese proposed terrain complexity features are combined with visual colour and texture descriptors and applied to classifying imagery. New multi-dataset feature selection methods are proposed for performing feature selection across multiple datasets, and are shown to improve the overall classification performance. The results show that the most informative predictors of benthic habitat types are the new terrain complexity measurements.
\nThis thesis presents a method that aims to reduce human labelling effort, while maximising classification performance by combining pre-clustering with active learning. The results support that utilising the structure of the unlabelled data in conjunction with uncertainty sampling can significantly reduce the number of labels required for a given level of accuracy.
\nTypically 0.00001–0.00007% of image data is annotated and processed for science purposes (20–50 points in 1–2% of the images). This thesis proposes a framework that uses existing human-annotated point labels to train a superpixel-based automated classification system, which can extrapolate the classified results to every pixel across all the images of an entire survey.