Research And Application Of Wildfire Visual Detection Based On Domain Adaptation

The wildfire,which occur in places such as forests,mountainous areas,and grasslands,is one of the natural disasters with strong suddenness,high destructive power,and difficult control,posing a serious threat to the ecological environment and personal safety.Therefore,how to detect and respond to wildfire incidents in a timely manner has become an important research topic.In recent years,with the development of computer vision and deep learning technology,wildfire visual detection technology has gradually been recognized for its advantages in fast speed,wide range,and low maintenance costs.Both image processingbased techniques and deep learning-based methods for visual detection of wildfires present numerous challenges in complex environments.On the one hand,wildfire objects do not have fixed shapes,making it difficult to design the feature extraction operators.On the other hand,high-quality wildfire image data is relatively scarce,and it is difficult for wildfire visual detection methods based on deep learning to learn enough feature knowledge from it.To overcome these problems,this thesis conducts research on wildfire visual detection methods based on domain adaptation.The main research content and innovation points of the thesis are as follows:(1)A fine-grained domain adaptation-based wildfire localization detection method is proposed in this thesis.The simulated wildfire images can assist the algorithm to learn wildfire features,but there are differences in feature distribution between the simulated images and the real images,and how to effectively overcome the discrepancy in feature distribution becomes the key.First,this thesis proposes a domain adaptive mutual attention mechanism,which measures the correlation between the simulated and real wildfire image feature vectors and guides the algorithm to prioritize the feature distribution with higher correlation to achieve fine-grained domain adaptation.Later,the fine-grained domain adaptive method is used to align the image-level and instance-level feature distributions of YOLOv5 to achieve fine-grained domain adaptive wildfire region localization detection.The experiments show that this method can effectively align the feature distributions of simulated and real wildfire images,improving the detection accuracy from 64.6% to 86.8%.(2)A domain adaptive wildfire region localization and segmentation joint detection method is proposed.Simultaneously conducting wildfire region localization and segmentation detection on a shared feature layer can further enhance localization detection ability.This thesis designs a dual-stream network for localization detection and segmentation detection,using the pixel information of the wildfire area extracted by the segmentation network to enhance the representation ability and improve the detection ability of the localization detection network.At the same time,it uses the instance features extracted by the localization detection network to strengthen semantic information and feedback to the segmentation detection ability.The experiments show that the proposed method effectively improves the performance of wildfire detection.(3)A domain adaptive wildfire visual detection system based on domain is implemented.In this thesis,algorithms designed above are integrated into the detection system according to the actual requirements of the wildfire visual detection task.And the system is tested to verify practicality.