Research On Forest Fire Smoke Recognition Technology Based On Video Images

In recent years,the situation of forest fire prevention in China has become more and more serious.Early identification and warning of forest fires are essential for fire fighting.This paper studies the forest smoke identification technology in the video surveillance scene.Because the forest smoke monitoring system covers a large scene,it is usually necessary to segment the suspected smoke areas first,and then perform smoke identification on these suspected areas.In view of the deployment environment of the practical application of the forest fireworks monitoring system,this paper studies the forest fireworks recognition technology based on low-speed front-end equipment and high-performance back-end equipment from the two key parts of the smoke area segmentation algorithm and the smoke recognition algorithm.The main content of the paper is as follows:(1)Fast smoke identification technology based on low-speed front-end equipment.When Vi Be motion detection algorithm is applied to the segmentation of smoke areas,there are problems that it cannot adapt to scenes with different depth of field,and the smoke segmentation is inaccurate and incomplete.In response to this problem,this paper proposes a Vi Be algorithm with adaptive depth of field.This algorithm calculates the transmittance of the scene based on the dark channel prior theory and roughly estimates the depth of field information.The Otsu threshold segmentation algorithm divides the image into near and far according to the depth.In different areas,different Vi Be motion detection sensitivity parameters are set respectively,which improves the adaptability of Vi Be algorithm to the depth of field.In the traditional smoke recognition algorithm,the smoke area segmentation is inaccurate,so that the smoke recognition algorithm based on image classification cannot accurately learn the characteristics of actual smoke,and the recognition accuracy is difficult to improve.To solve this problem,this paper expands the receptive field of the recognition algorithm,and uses the YOLO object detection algorithm based on deep learning for the smoke recognition problem.Aiming at the problem of large scale and slow execution speed of the network,an improved fast YOLO network is designed.The network uses a backbone network based on CBAM attention module and depth-wise separable convolution,which improves the feature extraction capability and execution speed,and prunes the number of scale branches mapped by the anchor box,reducing the size of the network and further improving network execution speed.(2)Smoke identification technology based on high-performance back-end equipment.This paper introduces a saliency detection algorithm based on a fully convolutional network.Through the pixel-level classification of the image,it overcomes the problem that the traditional motion detection method cannot segment the internal area of smoke.Aiming at the problem that the original full convolutional network smoke segmentation is not accurate enough and there are many false detections,a full convolutional network based on feature pyramid is designed.The network forms a multi-layer feature map through the hollow convolution of different receptive fields,and combines high-dimensional and low-dimensional features to form a feature pyramid structure,which enables the network to perceive more surrounding information and improve the network’s ability to adapt to different scales of smoke.In order to further improve the accuracy of the smoke recognition algorithm,this paper introduces Faster R-CNN network and optimizes the network recognition accuracy,designing an improved cascade R-CNN network.The network uses a backbone network combined with CBAM modules and aggregated residual learning.By cascading multiple detector networks with increasing Io U thresholds,the problem of large differences in the quality of candidate boxes during the training and forward calculation stages of the Faster R-CNN network is alleviated,which further improves the accuracy of the algorithm.(3)Design and implementation of forest fireworks identification software.According to the forest fireworks recognition algorithm studied in this paper,a forest fireworks recognition software is designed and implemented.The software implements the region segmentation algorithm based on the full convolutional network used on high-performance devices and the improved cascade R-CNN smoke recognition algorithm,as well as the adaptive depth-of-field Vi Be algorithm used on low-speed front-end devices and the improved fast YOLO algorithm smoke recognition algorithm.The software is tested on the video test set made up of actual scene.Experimental results show that the forest smoke identification software designed in this paper can satisfy the accuracy and real-time requirements of forest smoke identification on high-performance back-end equipment and low-speed front-end equipment.