Forest Fire Risk Image Recognition Based On DBN-CNN Network

This paper studies the smoke recognition based on DBN-CNN.The specific research contents are as follows:(1)The smoke generative adversarial network framework SMOKE-GAN is proposed to generate simulated forest fire smoke images and then expand the training sample data and avoid overfitting due to insufficient data.The smoke generative adversarial network framework SMOKE-GAN is composed of three networks: generator,discriminator and smoke dedicated network.To train a convolutional neural network to ensure that there are sufficient samples,the simulated synthetic smoke image data set obtained by the smoke generative adversarial network framework SMOKE-GAN and the real smoke image data set are combined to solve the problem of lack of smoke video data.The special smoke network in SMOKE-GAN helps to make the smoke image training process more standardized and solve the problem that traditional GAN is prone to confrontational losses.The experimental results show that the correct probability that the image generated by SMOKE-GAN is classified as smoke is92.35%.(2)GB-DBN is proposed to pre-process the sample data.Add Gaussian-Bernoulli RBM to traditional DBN for unsupervised layer-by-layer optimization.The fragmentation operation of the improved deep belief network discards the existing relationship between adjacent fragments and thus the redundancy between adjacent features can be kept at a low level;the interference information in the image is discarded(Noise,etc.),which in turn highlights the useful information in the picture.The feature information in the preprocessed image is more concise,efficient and independent,which is convenient for the initialization of the relevant parameters of the convolutional neural network.The experimental results show that the classified test loss rate of the GB-DBN algorithm model is 0.25%.(3)DBN-CNN is proposed to identify forest fire smoke.When using convolutional neural networks for smoke recognition with a small amount of data and rich content,too many layers will cause over-fitting,which will reduce the accuracy of detection.The deep belief convolutional neural network structure constructed in this paper show that the DBN-CNN algorithm model from the experimental results has an accuracy rate of 98.52% for forest fire smoke recognition.