| Wildfire is a sudden and destructive natural disaster.Once the best time to fight fire is missed,it will cause great economic and property losses.The introduction of computer vision technology for intelligent wildfire detection can help minimize property losses.However,due to the lack of image data for wildfires,the current wildfire detection algorithms generally have a high false alarm rate,which hinders development and application of computer vision technology in wildfire prevention scenarios.An effective way to solve this problem is to enhance the generalization performance of the model from the data perspective by expanding the wildfire image dataset with data augmentation methods.Based on the application of various generative models in data generation,this thesis does research on wildfire image data enhancement method combined with the actual background of wildfire detection.This research is of great significance for improving the accuracy of wildfire detection and reducing the false alarm rate of detection.The work of this thesis includes the following contents:(1)Proposing an approach for generating wildfire images based on a diffusion model.To solve the serious lack of existing wildfire images,this thesis proposes a method for generating wildfire images based on a diffusion model,which can generate wildfire images in batches that conform to the actual detection scene,and effectively expand the wildfire image dataset.This method combines the advantages of two generative models,the variational autoencoder and the diffusion model,applies the diffusion process to the latent space,and proposes a global attention residual module to assign different weights to each target in the wildfire image,so that typical features of small targets in the image have a better generation effect.Experiments show that the generated images obtained by this method have a high probability distribution similarity with real wildfire images,which can be used as training data to improve the detection accuracy of the wildfire detection model and reduce the false alarm rate.(2)Proposing an approach for translation of wildfire images based on generative adversarial networks.This thesis designs a wildfire image translation method based on generative adversarial network,which is used to convert ordinary wildfire images into different environments and different styles of wildfire images to expand the wildfire image dataset.Based on the recurrent generation of confrontation network,this method adds a global attention residual module to the generator to perform more refined image domain conversion on the typical features of wildfire images,keeping the data distribution of related image content unchanged;local discriminator is added and a diffusion model is introduced to enhance image translation.Experiments have proved that this method has achieved good results in quantitative analysis and comparison on multiple data sets,and can obtain better translation effects of mountain fire images.(3)Designing and achieving a wildfire image data enhancement system.This thesis analyzes the actual needs of wildfire image data enhancement,designs and implements a wildfire image data enhancement system,and integrates the proposed image generation algorithm and image translation algorithm into the system for industry application. |
