PROJECT TITLE :

A Wave-Shaped Deep Neural Network for Smoke Density Estimation

ABSTRACT:

Smoke density estimate from a single image is an entirely new but severely ill-posed problem.. W-Net is a wave-shaped neural network we developed to combine different convolutional encoder-decoder structures to tackle the problem. Increased receptive fields for storing more semantic information can be achieved by stacking encoder-decoders. Short-cut connections for spatial accuracy are achieved by concatenating outputs from encoders with those from decoders and copying the outputs and resizing them. Short-cut connections are used to connect the W-Net structures and decoding layers since these structures include a lot of localization and semantic information. Smoke segmentation, smoke detection, and disaster modelling all benefit from an estimated smoke density. Both smoke density estimate and segmentation can be improved by our method, according to the experimental results. Visual detection of automobile exhausts is another a successful application of this technology.