Mining enterprises production scene deployed a large number of monitoring equipment,using the object detection technology specific object real-time monitoring image object detection in mines,helps to establish a safe and efficient production environment.The deep object detection model with large number of parameters and computation is difficult to deploy to the edge end to complete the real-time Mining object detection task.The detection accuracy of the lightweight object detection model is affected by factors such as image clarity and training sample data amount,and the Mining object detection task is faced with the problem that the speed and accuracy cannot be improved simultaneously.In order to solve the above problems,this thesis established a mining object detection data set including miners,safety hats,coal gangue and other important objects in coal mine scenes,and conducted the following studies based on edge computing and federal learning technology:(1)In view of the problem that the deep object detection model relies on high-performance computing hardware and the high load of the central computing node makes it difficult to improve the real-time performance in practical application,this thesis designs a object detection method suitable for the object detection task of the edge end mine monitoring image.The coal mine video surveillance image data is collected to construct the Mining Object detection data set,which is used to evaluate the actual application performance of the object detection model.Simulating the experimental environment of edge Mining object detection,the designed YOLO-v4-Mobile Net-v3 model improves the m AP value of detection accuracy index by 1.2% relative to YOLO-v4-Tiny,and reduces the number of parameters and computation to about 1/6 and 1/9 of YOLO-v4 model,respectively.Can be deployed in mine monitoring equipment closer the edge of the end.(2)Aiming at the problem that the unequal illumination factors in the coal mine scene affect the image clarity and the object detection accuracy is low,this thesis proposes a lightweight Mining object detection method based on spatial attention.A spatial attention mechanism was added in the key layer of the backbone network to enhance the significant regional features and suppress the non-significant regional features in the spatial dimension.The pixel regularization spatial attention module(PNSAM)was proposed to enhance the attention of the model.Mish activation functions are used in the deep structure of backbone network to reduce gradient loss and retain richer gradient information.YOLO-v4L-SA model proposed in this thesis the experimental results on the demo dataset VOC2012 relative to YOLO-v4-Tiny object detection precision index m AP value increased by 3.2%,illustrates the effectiveness of the proposed model in common object detection scene.Compared with the object detection accuracy index m AP value of YOLO-v4-Tiny model,the experimental results in the Mining object detection data set are improved by 2.4%,indicating the effectiveness of the proposed model in coal mine scenarios.In the uneven illumination of the image in this thesis,the test results of the picture model error checking and less residual situation,this article put forward the model features extraction phase diagram of the corresponding region is more concentrated in the object area to be detected,that YOLO-v4L-SA coal mines in the uneven illumination of the image object detection is more effective.(3)Aiming at the problem that the Mining object detection model has low accuracy in terms of insufficient training sample data categories,this thesis proposes a Mining object detection data sharing scheme based on federated learning.In this thesis,the coalmine scene image data is further collected to expand the data set and the federal data set of Mining object detection is constructed.A federated learning scheme for Mining object detection was designed based on the aggregation algorithm of YOLO-v4L-SA model and Fed Avg federated learning model.In the experimental results,compared with the model precision index m AP value of chapter 3,the federal learning model training scheme improved by 2.5%,and the precision index m AP value of coal gangue category improved by 11.8%,indicating that the federal learning scheme improved the accuracy of Mining object detection model by expanding training data samples.Compared with the training results of centralized model with all the data,the model accuracy obtained by federal learning lost 3.9%,indicating that the Mining object detection scheme of federal learning in this thesis still has much room for improvement.(4)Based on the Mining object detection structure based on edge computing proposed in this thesis,a real-time object detection system for mine production scenes is designed and developed.It is mainly divided into user management module,real-time monitoring module and detection and record module,which can be used for users with different roles and multiple devices to log in. |