A deep convolution neural network fusing of color feature and spatio-temporal feature for smoke detection

Mengqi Ye, Yanmin Luo

Abstract: The spatial characteristics, movement characteristics and color characteristics of smoke are important features that distinguish them to other objects. In order to make full use of these three features, we proposed a deep convolutional network called Full High Resolution Network(FHRNet).This network consists of two parts: Spatio-Temporal-aware Sub-network (STS) and Color-aware Sub-network (CS). We build high -resolution residual symmetrical units and embed the two sub-networks to ensure the integrity of two dimensional features.In the STS, the residual symmetrical unit extracts the spatial semantic characteristics of smoke from every frame, and combine them into a feature sequence, then the spatio-temporal perceptron is used to extract the spatio-temporal characteristics of smoke to further improve the characteristic expression. In the CS, the color feature of picture is converted into color feature matrix, which is easier to make the residual symmetrical unit to extract the color feature of smoke. We constructed a smoke vedio datasets which have a diverse background to avoid producing over-fitting situation.The experimental results show that mthod we proposed can effectively extract the color features and the spatio-temporal features of smoke and our method can effectively detect smoke.

External IDs:dblp:journals/mta/YeL24