Detection Algorithm Of Smoke And Fire Objects Around Transmission Lines And Its Lightweight Research

In recent years,with China’s increasing dependence on electric energy and the continuous expansion of power grid scale,it has become crucial to ensure the safe operation of transmission lines.Fire in the surrounding environment of the transmission line will directly affect the normal operation of the transmission line and the normal power consumption of people’s life and production.It is necessary to detect the smoke and fire around the transmission line in time to find out the fire as soon as possible and reduce the harm of the fire.The quality of smoke and fire images collected around the transmission line is easily affected by the environment,and the amount of data is relatively scarce.In addition,the existing deep learning algorithms have insufficient ability to extract the features of smoke and fire targets with multiple scales and irregular shapes,resulting in low detection accuracy for smoke and fire targets.The detection algorithm based on deep learning also has the problem of large amount of parameters and computation,which makes it difficult to deploy and run in edge devices.In view of the above problems,this thesis studies the smoke and fire target detection algorithm around the transmission line and deploys the algorithm on the server and the edge respectively.The main work is as follows:(1)Smoke and fire image processing and data enhancement to improve image quality and enrich data diversity.In order to solve the problem of low contrast and blurred images of smoke and fire images in foggy weather,a contrast-limited adaptive histogram equalization dehazing algorithm and Laplacian image enhancement algorithm are used to improve the image quality.The K-means clustering algorithm is used to analyze the data of the constructed smoke and fire data set,and the anchor setting of the deep learning algorithm is improved to make the algorithm anchor better match the real smoke and fire target size.In order to solve the problem of lack of data,according to the characteristics of smoke and fire images,the data set is enhanced online from the aspects of image brightness,target size and position,image background and target number to enrich the data diversity.(2)The YOLOv4 algorithm is improved for the smoke and fire targets around the transmission line,and a smoke and fire detection algorithm based on feature enhancement and assisted positioning detection suitable for server deployment is designed.In order to adapt the algorithm to the irregular shape of the smoke and fire target and better extract the target features,the deformable convolution is used to improve the CSP feature extraction structure of the backbone network,and the shape of the convolution kernel is adaptively adjusted.In order to improve the feature extraction of multi-scale smoke and fire targets,an improved structure based on dual attention and multi receptive field feature enhancement fusion is designed.Dual attention enhancement algorithm is introduced to focus on target features and spatial positions.The multi-scale receptive field feature enhancement structure is used to fuse smoke and fire feature maps with different receptive field sizes.At the same time,a horizontal jumper structure is added to fuse the features of the same layer.In order to improve the positioning ability of the algorithm for smoke and fire targets,the detection head is improved,and the detection head structure based on auxiliary positioning is designed.Finally,Focal Loss is introduced to improve the loss function in order to solve the imbalance problem of difficult and easy samples in the training process of the algorithm.(3)In order to reduce the amount of parameters and computation of the algorithm,and make the smoke and fire detection algorithm more suitable for deployment in edge devices,this thesis studies three lightweight schemes: lightweight network structure design,network pruning and model quantization.In lightweight network structure design,lightweight network and lightweight convolution structure are used to improve the backbone network of the algorithm and enhance the feature extraction structure to achieve fine simplification of network structure.In the aspect of network pruning,the redundant convolution cores with high similarity in the network are pruned,and the similarity of convolution cores is introduced to improve the network pruning effect.In the aspect of model quantization,the quantization method is improved based on KL divergence and bias correction to reduce the precision loss in model quantization.(4)Design and realization of smoke and fire detection system.According to the task requirements,smoke and fire detection algorithms in this thesis are deployed to realize the smoke and fire detection function,and a server-based smoke and fire detection system and an edge-based smoke and fire detection system are designed respectively.