A Dual Convolution Network Using Dark Channel Prior for Image Smoke Classification

Abstract

Through a comparative analysis, we confirm that the value of the dark channel pixels of the smoke image is higher than the non-smoke image. It means that the dark channel of the smoke image has more elaborate information of the smoke, which is of great benefit to our detailed feature extraction of smoke. On this background, we propose a dual convolution network using dark channel prior for image smoke classification (DarkC-DCN) for the image smoke classification. In DarkC-DCN, basing on the AlexNet, and through continuous structural improvement and optimization, we improve a detailed CNN to extract the detailed features of dark channel images. Similarly, to extract the general features in the image, we further design another residual network based on the AlexNet, which is the main framework of the entire network. To ascertain the robustness of the network, the two channels are trained separately for various inputs. In addition, we perform feature fusion before the common fully connected layer. In the experiment, we also add some non-smoke data similar to smoke in the public smoke data set for data expansion. The experimental results indicate that the model has a good performance in general. The accuracy value reaches 98.56%.