The problem was part of the machine learning competition in inter IIT tech meet, 2018, where we won the silver medal in the competition. The task was to classify all pixels in a given satellite image, as belonging to one of nine given classes. Our approach towards solving this challenge involved three approaches, with each approach performing better than the last one. The first model involved the implementation of a basic U-Net architecture with custom metrics designed for this problem. The second approach involved a P-Net architecture with extensive hyperparameter tuning to improve model accuracy. The third approach utilizes nine distinct U-Net architectures for the segmentation, giving results comparable to the state of the art model for multi-class image segmentation.
My contributions to the project :
I designed and implemented a U-Net architecture for image segmentation of high-quality satellite images by using context-based representations.
I improved the existing accuracy from 84% to 91% on just 25 images by developing a new algorithm based on 9 U-Nets using ‘one vs all’ classification approach.
I used localized optimization of parameters with high frequency to break the bottleneck of a small dataset.
