Dark convolutional neural network for forest smoke detection and localization based on single image

Forest fire causes severe economic and ecological damage each year. Timely detection of forest fire in the early stage could effectively avoid possible disaster which requires efficient smoke recognition instead of fire identification. However, the appearance of smoke might share similarity with the background which makes accurate smoke detection difficult. In addition, in contrary to the normal state forest images, it is much more difficult to collect early stage forest fire images, so the available training data for forest smoke detection is severely imbalanced. To address the above issues, a dark channel boosted convolutional neural network has been developed for smoke detection termed as DarkCNN. A Min pooling layer is designed and adopted next to the input layer to extract the dark channel feature of the image. The dark channel of an image could enhance the difference between the smoke and non-smoke pixels and thus provide more discriminant features. To address the imbalanced data distribution problem, weighted softmax loss was adopted to train the DarkCNN. In addition, to localize the smoke region within an image with low time cost, a pipeline combining Simple Linear Iterative Cluster (SLIC) algorithm and DarkCNN has been developed which has much lower computational complexity than methods like R-CNN and its variants. SLIC is employed to segment the image and generate smoke candidates for DarkCNN. The classification results are used to localize the smoke region. Extensive experiments have verified the efficiency of DarkCNN for smoke detection based on single frame.