Convolutional neural networks based on multi-scale additive merging layers for visual smoke recognition

摘要

Traditional smoke recognition methods are mainly based on handcrafted features. However, it is difficult to design handcrafted features that are robust and discriminative for smoke recognition because of large variations in smoke color, shapes and textures. To solve this problem, we specifically design a basic block of convolutional neural networks (CNNs) and stack basic blocks to propose a novel deep multi-scale CNN (DMCNN) for smoke recognition. The basic block consists of several parallel convolutional layers with the same number of filters but different kernel sizes for scale invariance. Each convolutional layer is followed by a batch normalization to normalize the output of the convolutional layer. Then the basic block sums up all normalized outputs from multi-scale parallel layers and activates the sum as the final output of the block. To fully extract scale invariant features, we cascade eleven basic blocks, which is followed by a global average pooling and a 2D fully connected layer, to construct DMCNN. Experimental results show that our method achieves higher detection rates, higher accuracy rates and lower false alarm rates than existing methods. To further verify the efficiency of DMCNN, we also conducted face gender recognition experiments on the LFW database and our model also achieves obviously higher accuracy rates than other methods. Furthermore, our method is an efficient, lightweight CNN model with about 1 M parameters that are far less than other CNN methods.