Efficient 2D image segmentation

Image segmentation is a popular topic enabled by rapid advances in neural network processing. There is also a trend towards computational efficiency in segmentation algorithms, brought about by economic and environmental interests, and often acutely motivated by constraints in how the solutions are deployed. Recent exploration of neural network computational efficiency has largely focused on the depth and width of the network, as well as image size. Ensembling and gradient boosting are two well-known methods for increasing performance through a collection of smaller networks. Ensembling is used when different models produce largely uncorrelated predictions; the combination of these predictions reduces overall error. Gradient boosting produces correlated models such that the sum of their losses is minimized; successive models are trained using the residual loss of previous models, with each new learner correcting mistakes of past learners. This thesis seeks to explore the scalability and efficiency of image segmentation neural networks by including the techniques of ensembling and gradient boosting. The approach is evaluated on both the Severstal Steel Defect Detection and Kidney Tumor Segmentation datasets, though it could easily be applied to any two-dimensional image segmentation task. The final boosted model produces results comparable to the baseline in 5 of the 6 evaluated classes, while using only 2.5% of the trainable parameters. The success of the approach has significant implications for deployed applications; not only is the overall network much smaller, the miniature sub-networks can each be calculated independently and aggregated with simple addition and averaging. This introduces a real possibility for efficient distributed inferencing, decoupling performance from the maximum size calculable by a single piece of hardware.Science, Irving K. Barber Faculty of (Okanagan)Computer Science, Mathematics, Physics and Statistics, Department of (Okanagan)Graduat