Fire Smoke Detection Method Based On Deep Neural Network

With the rapid development of deep learning in recent years,many scholars have combined traditional computer research with deep learning techniques to solve practical problems in real life,and have obtained rich research results.Among them,the combination of deep learning and target detection has attracted the attention of more and more researchers.Progress and increasingly low prices make the computing power required to train deep learning models easy to acquire,and with the advent of the “big data” era,massive data becomes easier to centrally manage and organize into data sets that are easy to learn in depth.Further promote the application and progress of deep learning in engineering technology and academic research.Smoke detection can play a key role in early warning of fires,while traditional smoke detection methods rely on professional detection equipment,sensitivity to environmental changes,and poor robustness.Therefore,this paper builds a deep neural network based smoke detection method based on the existing deep learning target detection technology.In this paper,an improved VGG16 network model is proposed.The smoke characteristics are extracted by convolutional neural network(CNN).The image background is complex and the captured image contains a lot of noise.The original VGG16 model has a poor effect on smoke feature extraction.In order to improve the convolution network.The accuracy of smoke feature extraction reduces the irrelevant features in the final feature map.This paper visualizes the output of each layer of the convolutional network to extract the intermediate layer features with obvious smoke.By comparing the visualization results of multiple scenes and multiple smoke intermediate layers,three layers are selected as the feature map output.Through experimental comparison,the improved VGG16 model proposed in this paper has a 3.5%improvement in classification accuracy of smoke images compared with the original VGG16 model.Problems in the smoke detection for the original target detection model.In this paper,we propose the conditional generation confrontation network(CGAN)with the smoke feature map as the constraint input.In order to improve the effect of generating the high frequency part(boundary)of the image generated by the network generator,the generator adopts the U-Net model.Discriminator is a PatchGAN structure.Under the same data set,the average IOU value of the smoke detection model is improved by 9%,5%,14%and average detection compared with Faster R-CNN,RetinaNet,and MobileNet.Faster than Faster R-CNN and RetinaNet.Have a certain effect advantage.