Additive neural network for forest fire detection

Abstract

In this paper, we introduce a video-based wildfire detection scheme based on a computationally efficient additive deep neural network, which we call AddNet. This AddNet is based on a multiplication-free vector operator, which performs only addition and sign manipulation operations. In this regard, we construct a dot product-like operation from the mf-operator and use it to define dense and convolutional feed-forwarding passes in AddNet. We train AddNet on images taken from forestry surveillance cameras. Our experiments show that AddNet can achieve a time-saving by 12.4% when compared to an equivalent regular convolutional neural network (CNN). Furthermore, the smoke recognition performance of AddNet is as good as regular CNNs and substantially better than binary-weight neural networks.

Computationally Efficient Wildfire Detection Method Using a Deep Convolutional Network Pruned via Fourier Analysis

“…We use the dataset of Reference [28] which contains 4000 images gathered from the Internet (half with fire and half without fire) shown in Figure 1a,b, and FIRESENSE database [37] shown in Figure 1c,d which contains 11 fire videos and 16 no-fire videos. During practice, we notice that the system would false alarm at cloud region sometimes, so we also add 4 cloud video clips shown in Figure 1e,f into our training dataset to reduce the false-alarm rate.…”

“…All images and frames are resized into 224 × 224 as the input of the neural network, and 5% of them are picked randomly for validation. Data augmentation by shifting and translating wildfire images with filling by reflect translation as in Reference [28] is also adopted here. Let Y ∈ R M×N denote the original image and index start from 0, if we want to shift the image by m length up and n length left (negative value means shifting in the opposite direction), then the augmented imageỸ can be represent as…”

“…To the best of our knowledge, Gunay et al [23] is the first paper to use deep learning in dynamic texture recognition including wildfires. In early deep learning based forest detection, researchers designed blank convolutional neural networks and trained them with collected or synthesized images [24][25][26][27][28]. For example, Zhao et al designed a 15-layer convolutional neural network (CNN) to detect the forest fire [25].…”

“…Compared with other deep learning methods [23][24][25][26][27][28][29][30][31][33][34][35], the major advantage is that, after transfer learning, we prune the network via Fourier analysis. More specifically, we have the following advantages:…”

“…Our system can detect very small smoke regions, while papers from related works [24][25][26][27][28][29][30][31] have not provided the test results on such small smoke regions. The details of this aspect are provided in Section 5.2.…”