Smoke And Fire Detection Algorithm Based On Deep Learning

The occurrence of fire will lead to ecological damage,property loss and casualties.In order to reduce this kind of disaster,it is very important to apply the video monitoring method to detect the fire and smoke objects and alarm automatically.However,the existing algorithms for fire and smoke detection still have some problems such as false detection,missed detection,difficulties in small objects detection and deployment for edge devices.To solve these problems,the light-weighted algorithm for fire and smoke detection based on deep learning are researched in this dissertation.The main work is as follows:(1)Aiming at the problem of false detection and missed detection caused by complex background and fire-like or smoke-like objects in existing fire and smoke detection algorithms,a new light-weighted fire and smoke detection algorithm i.e.EGC-YOLOX based on Efficient Global Context Network(EGC-Net)is proposed.It takes the light-weighted object detection network YOLOX as the basic network,and embeds an improved EGC-Net between the backbone feature extraction network and feature pyramid network of YOLOX.EGC-Net is composed of a three-stage structure of context modeling,feature transformation and feature fusion,which is used to obtain the global context information of the image,model the long-range dependency of the pyrotechnic object and its background information,and combine the channel attention mechanism to learn more discriminative visual features used for fire and smoke detection.The experimental results show that the image-level recall rate of the method for fire and smoke images in the test set is 95.56%,and the image-level false alarm rate is 4.75%,both of which are better than other compared typical lightweight algorithms,and the speed meets the requirements of real-time detection.(2)Aiming at the problem that existing algorithms are difficult to detect small objects of fire and smoke,an improved EGC-YOLOX algorithm for fire and smoke detection based on Small Object Augmentation(SOA)and Balanced Loss Function(BLF)is proposed.SOA obtains samples containing small objects of fire and smoke by adaptively scaling the image size,and embeds them in random positions on randomly selected images containing fire-like or smoke-like objects,and updates the position labels of small objects of fire and smoke accordingly,thereby augments the small object data of fire and smoke.In addition,the BLF loss function is proposed to balance the regression loss of small objects and normal objects,and improve the small object detection ability of the algorithm.The experimental results show that the small object detection rate of the algorithm reaches 96.56%,while the image-level recall rate reaches97.52%,and the image-level false positive rate is only 2.36%,which is better than other compared algorithms.(3)Aiming at the application requirements of deploying the fire and smoke detection algorithm on embedded devices,the model compression of the proposed fire and smoke detection algorithm is carried out,then an ultra-lightweight model is obtained and deployed on the Jestson Nano development board through structured pruning and quantization.The effectiveness of the model compression method and demonstrates the operation effect of the proposed smoke and fire detection method.