Machine Learning Approach for Adaptive Automated Smoke Region Detection

Automatically detecting smoke and fire in natural scenes can potentially lead to forest fire prevention and therefore to saving millions of budget money.In this thesis, we develop a novel algorithm that detects potential smoke regions in images.Data comes in the form of sequences of three images (image triplets) of the each scene with a bounding box highlighting the smoke whereabouts.The algorithm extracts features from these images and uses unsupervised machine learning techniques to detect potential smoke regions.First, The algorithm divides images into superpixels, extracts features per superpixel and then clusters the superpixels in such a way that separates smoke regions from non-smoke ones.The novelty of the algorithm comes from the fact that it operates in a semi-supervised manner due to the absence of exact labels.It is built in a modular way that is easy to debug and extend.It doesn't rely on fixed thresholds which makes it independent of specific datasets and thus can generalize to new, unseen conditions.i Acknowledgement I am heartily thankful to all those who