A Wave-shaped Deep Neural Network for Smoke Density Estimation

Abstract

Smoke density estimation from a single image is a totally new but highly ill-posed problem. To solve the problem, we stack several convolutional encoder-decoder structures together to propose a wave-shaped neural network, termed W-Net. Stacking encoder-decoders directly increases the network depth, leading to the enlargement of receptive fields for encoding more semantic information. To maximize the degrees of feature re-usage, we copy and resize the outputs of encoding layers to corresponding decoding layers, and then concatenate them to implement shortcut connections for improving spatial accuracy. The crests and troughs of W-Net are special structures containing abundant localization and semantic information, so we also use short-cut connections between these structures and decoding layers. Estimated smoke density is useful in many applications, such as smoke segmentation, smoke detection, disaster simulation.

Existing System

Spatial pyramid pooling (SPP), R-CNN

Proposed System

The main contributions of this paper are as follows: 1) We propose an end-to-end smoke density estimation method using full convolutional networks (FCN). The proposed deep network can be trained end-to-end and used to estimate a smoke density map from an observed image of arbitrary size. 2) We stack several non-symmetric encoder-decoder networks together to propose wave-shaped structures, which greatly expands receptive fields of each neuron, and forms special wave structures, including troughs and wrests. The troughs of wave structures consist of abundant semantic information for dense classification over all pixels, while the crests of wave structures contain more local or median-scale information for smoke localization. To improve accuracy, we use short-cut connections between troughs, crests and decoding layers. 3) We propose a complicated loss function that includes four errors of smoke density, smoke color, background color, and composited observed color. These four-error terms jointly regulate the training of our proposed network.

Architecture

BLOCK DIAGRAM