Research And Application Of Efficient Flame Recognition Algorithm Based On Deep Learning

In fire monitoring and early warning,early detection and early response are crucial to controlling the development of fires.The use of flame recognition technology can help improve detection speed and accuracy,and win valuable time for reducing fire losses.Therefore,studying deep learning algorithms for flame recognition has theoretical significance and application value.The work done in this article is:(1)Taking the identification of gunpowder flame as the research object,a learning model structure is designed to maintain fast reasoning speed and high model accuracy.To ensure real-time performance,the lightweight neural network ShuffleNetV2 is selected as the backbone network;In order to improve the accuracy of the model,an attention module(space and channel dual attention module,SCDAM)was designed considering channel and spatial correlations to perform weight management based on the importance of different features;To enrich the extracted features on spatial scales,a multi scale feature fusion module is designed to enhance the adaptability of the network to different scales;After introducing the SCDAM module and multiscale module into ShuffleNetV2,transfer learning is used to optimize model parameters and further improve model accuracy.The experimental results show that under the condition that the number of parameters and computation amount increase is limited,the accuracy of the algorithm is improved by 3.2% compared to ShuffleNetV2,and the single inference time reaches 8.7 ms,which can meet the application requirements under the condition of limited computing resources.(2)A lightweight pyramid hybrid architecture network model,SconvTrans,is designed to solve the problem of fire like image interference in the process of flame recognition.To ensure accuracy while reducing the amount of computation and achieving lightweight model,ShuffleNetV2’s downsampling block is used to downsample the input image,and Transformer and convolution blocks are used for computation;To enhance the model’s ability to extract and express features,a CTB module integrating Transformer and Convolutional Neural Network was designed to fuse local information features extracted by Convolutional Neural Network and global information features extracted by Transformer.Experimental results show that under the condition of lightweight model,it can maintain high model accuracy and effectively identify fire like image interference.(3)A flame recognition monitoring system based on deep learning was designed and developed,mainly including a real-time monitoring module,a flame recognition module,an alarm module,and a data storage and management module.The test results show that the system has the function of fire monitoring and warning in specific environments.