Lightweight Real-time Semantic Segmentation And Its Application In Smoke Detection

Fire is one of the most common,high frequency,and more harmful disasters in human society.Although the sensor-based smoke detection is currently a widely used and low-cost fire waring method,it does not meet the current high requirements of detection range,application environment and time delay.With the development of computer vision and artificial intelligence,smoke detection technology based on video images has the advantages of wide applicability,high real-time performance,safety and reliability,and multiple functions,and has gradually attracted people’s attention.Due to the diversified color,shape,concentration,and situation of the smoke images,the generalization ability of the smoke detection model based on the traditional smoke image characteristics is weak,the prediction accuracy of the smoke is low,and it is easy to falsely report and omit the report.The current smoke detection methods based on deep learning effectively improve the accuracy of smoke recognition.However,the model complexity is high and can’t be adapted to mobile terminals easily.Based on the above problems,the main work of this article is as follows:(1)In view of the problem of weak feature extraction ability and high parameter quantity of existing convolution modules,an enhanced asymmetric convolution module is proposed based on asymmetric convolution,deep convolution and decomposition convolution,which can effectively capture local and global features information.Meanwhile,the blocks are efficient and lightweight In view of the shortcomings of conventional semantic segmentation networks that only focus on the semantic information and ignore the spatial detail information,a lightweight real-time semantic segmentation network is proposed based on the enhanced asymmetric convolution modules.The network adopts a dual-branch structure to extract high-level context information and low-level spatial detail information respectively.At the same time,the two branches are merged multiple times in the process of feature extraction to enhance the feature information interaction and propagation of the network and effectively improve the ability to capture the characteristics of smoke.(2)Aiming at the problems of low accuracy of existing smoke algorithms,slow inference speed,and imprecise boundary segmentation of existing smoke algorithms,a smoke detection algorithm based on lightweight real-time semantic segmentation is proposed.Due to the uneven concentration of the smoke edge,the edge of the smoke area is fuzzy,which cannot be accurately predicted and segmented.Therefore,a multi-level fusion optimization module and an efficient coordination attention module are added to the lightweight real-time semantic segmentation network.A series of experiments have proved that the network has high performance and efficiency,which provide theoretical support for the deployment of algorithms to the mobile terminal.