Result: Machine learning-based segmentation applied to wind turbines images: loss exploration

This collaborative study with Wind Power Lab and IRI aimed to enhance image segmentation algorithms, specifically region growing, through innovative image pre-processing techniques. The core objective was to modify the color space of images using statistical methods and optimization tools to improve algorithm performance. The project involved the development and analysis of the region growing algorithm. We explored various colour scales and assessed their impact on segmentation outcomes, leading to a deeper understanding of customised image colour scale modifications. The pivotal phase of this study involved formulating an optimization problem that targeted a linear constant transformation of image colours. The aim was to improve windmill identification against varied backgrounds. This was achieved by minimizing the distances between pixels associated with windmills while maximizing background pixel distances. The approach used was statistically analogous to binary classification problems. Analytical and experimental methods, including gradient descent, were used to define a linear transformer vector for modifying colour channels. Despite challenges in transforming colour space into three dimensions and limited testing data, the algorithm demonstrated improved segmentation in three dimensions. This suggests that it is effective in preprocessing before segmentation. The study concluded that while colour space transformation can enhance image segmentation, the seeded region growing algorithm showed superior results.