Research On Smoke Detection Based On Light-weighted Convolutional Neural Network

Smoke detection,as a prior technical means of fire warning,is very important for modern intelligent fire protection system.However,the scene of smoke image is fuzzy,and the background is complex and changeable.With the rapid development of computer vision and pattern recognition technology,smoke detection and recognition algorithms based on deep learning have emerged in recent years.Compared with the traditional point-based sensor early warning system,it has made a breakthrough in the accuracy and application range,providing reliable technical support for fire early warning.Although the current smoke detection algorithms have made some achievements in accuracy,it is not optimistic that most of them have such problems as high computational complexity,long time consumption,high false alarm rate and missing alarm rate.In this paper,a series of exploration and research on the optimization of smoke detection algorithm are carried out.This paper proposes an end-to-end lightweight convolutional network,named SInception.Firstly,the network adjusts the Inception structure of GoogLeNet,so as to improve the accuracy of smoke detection while retaining the characteristics of low computation.Then Sequeeze-and-Excitation block is used to combine with adjusted SInception structure.The channel weights in the feature graph are automatically allocated to help the network improve the ability of feature expression.In this paper,the data set YUAN is used for training.The recognition accuracy of this algorithm reaches 98.68%,the detection rate is 96.68%,the false detection rate is 0,and the detection speed is only 0.2ms.Comparing with the current most advanced smoke detection algorithm DNCNN,this algorithm improves the detection rate by 1.11%,the number of network parameters is only8.8%,and the computing speed is improved by 7 times.In order to make efficient use of the RGB features of smoke image and dark channel features,and improve the accuracy of the algorithm,this paper first extracts dark channel data from the original data set to make dark channel data set,and uses the data set to train the proposed network to obtain dark channel prior model.Then,the probability fusion strategy and Max fusion function are adopted to fuse the predicted results of the standard model and the dark channel model without destroying the pre-training model.The recognition accuracy of fusion algorithm is improved to 99.07%.This paper adopts data augmentation and deep convolution generative adversarial network to explore the impact of data distribution on algorithm performance.Using the opendata set YUAN to generate more training samples,and mixing the original data and generated data to make the extended data set GAN-AUG-YUAN.In this paper,a large number of comparative experiments are carried out on the original data set and the extended data set,in which the recognition accuracy of the fusion model is as high as 99.79%,surpassing all previous smoke recognition algorithms.The experimental results show that in the current state of smoke detection accuracy approaching saturation,data expansion can effectively improve the accuracy of model detection.