Research On Visual Fire Smoke Detection Method For Large Space Building

Visual fire detection in large-space buildings is an important method of using computer vision technology to analyze and process images of buildings with large areas,large spans and high storeys so as to determine whether a fire has occurred.Compared with traditional fire detection methods such as temperature and smoke sensing,it has the advantages of large monitoring range and robustness,and is suitable for fire warning in large space buildings(such as industrial plants,large shopping malls,etc.).When a fire occurs,smoke often precedes flame,a feature that is important for early identification of fire.Smoke is a non-rigid flexible target,and there are many problems for its target detection,such as suspected smoke misdetection and difficulty in detecting smaller smoke targets.To address the above problems,the work carried out in this paper is as follows:(1)A fire smoke segmentation and recognition method based on density peak clustering is proposed for the problem that inconspicuous smoke boundaries and weak light smoke features can lead to poor recognition results.First,the image is segmented into multiple superpixels by using the superpixel technique,and the local density of the sampling points corresponding to each superpixel is redefined by the position information and color space information;second,the optimal double truncation distance is found by combining the information entropy theory;finally,the knowledge of trigonometric function is used to determine the number of clustering centers in the decision diagram,and the remaining sample points are assigned to complete the extraction of the suspected smoke image;on the Based on the segmentation results,the fire smoke image recognition is completed by combining SVM.Experimental validation using self-built and public datasets shows that the method can better extract the target smoke regions in the suspected fire smoke images and improve the accuracy of smoke recognition.(2)A fire smoke detection method based on depth-separable convolution is proposed to address the problem of small smoke targets at the early stage of fire and the difficulty of timely detection.Based on a single-stage target detection model,the method uses Mobile Net as the backbone feature extraction network,extracts more image features from the shallow feature map through the convolutional layer,enhances the key feature information of the image through the channel attention module,and then passes the feature map through the light-weighted sensory field expansion module to enhance the adaptability of the network to the multi-scale changes of smoke targets.In addition,an effective channel attention mechanism module is added to improve the detection performance of the convolutional network.The experimental results show that the method can detect smoke targets in fire images in real time and has high detection accuracy.