Federated Vision And Its Usage In Air Pollution Monitoring

With the development of industrialization,air pollution monitoring is gradually attracting attention.In recent years,researchers are gradually focusing on the analysis of air pollution based on computer vision technology.In order to perform real-time detection of industrial smoke and guarantee data privacy during model training,this paper proposes a factory smoke detection algorithm combined with channel attention mechanism based on federated learning technology,through which the training of smoke detection models can be achieved in a privacy-preserving environment,and can also be used for industrial smoke emission detection in environmental protection departments.In the background of big data,the theory of deep learning,which supports training on large amounts of data,has emerged.Deep learning-based target detection methods can perform detection of a large number of smoke images more effectively than traditional image processing methods.Since in real scenarios of smoke detection,in most cases,the detection targets are distributed over a large area and are easily blended with the background,problems such as missed and false smoke detection exist from time to time.In addition,the prevalent data barriers among enterprises also limit data interoperability and affect the feasibility of training large-scale deep learning models.In this paper,a series of studies have been conducted to address the above issues and the main contexts are as follows:(1)Based on YOLO-v5 object detection framework,combining with the channel attention mechanism,this paper proposes an optimization algorithm,YOLO-v5s-SE,to filter the channels through the SE block structure.This algorithm gives more weight to the target,and effectively attenuates the influence of background on target detection.(2)In this paper,based on the existing smoke detection dataset,a federated smoke dataset is formed according to different sources and scenarios,and divided into different clients for the evaluation and analysis of the used target detection methods.(3)This paper is tested in both centralized training and federated learning environments.The accuracy of the proposed model is better than the benchmark model in the centralized training environment.In the federated learning training environment,the effect of the proposed model is slightly worse than that in the centralized environment,but it is comparable to that of the YOLO-v5s benchmark model.In addition,in terms of detection speed,the YOLO-v5s-SE model can reach 34 ms,which can meet the real-time requirements and can be applied to the devices at the federated learning end,and can also basically meet the industrial real-time requirements.In summary,the channel attention mechanism-based target detection YOLOv5s-SE algorithm proposed in this paper can be applied to the smoke detection task and can be trained with data privacy protection,and can be applied to the smoke detection scenario in the federated learning scenario,which has some social significance.