Water Level Detection Based On Faster R-CNN And GrabCut

Water level monitoring of rivers,reservoirs,lakes and other drainages is a key link to prevent flood disasters and drought.With the development of video monitoring technology and network communication technology,the use of video monitoring system to monitor water level is increasingly popular.The water level detection point usually were setted up in River area by the Water conservancy department,installs water gauges and cameras,artificial observation to obtain water level through video monitoring.The task of this method is difficult to achieve real-time performance.It is urgent to study the real-time,automatic and accurate water level monitoring method and technology based on video image.This thesis designs an automatic water level detection algorithm based on the fusion of Faster R-CNN and Grab Cut.First,the Faster R-CNN network was used to detect the water surface above the draft,then use the Grab Cut algorithm to accurately segmentation,and finally calculate the water level value using the mapping relationship between pixel coordinates and world coordinates.In order to improve the positioning accuracy of Faster R-CNN network and the segmentation efficiency of Grab Cut algorithm,this paper mainly includes the following four aspects.(1)Part of the parameters and structure of Faster R-CNN network mainly designed for the data set VOC2007.In order to improve the sensitivity of Faster R-CNN network to draft target,feature fusion is introduced into feature extraction network.After up-sampling or down-sampling,different receptive field feature maps are fused to produce multi-scale feature maps with rich feature information and stronger robustness.In the RPN network,K-means clustering algorithm is used to initialize the Anchor frame proportion,which makes it more consistent with the characteristics of the draft target,reduces the difficulty of border regression,and improves the positioning accuracy of the draft.In order to improve the speed of target detection in Faster R-CNN network,the part of 3×3 convolution layer in VGG16 network model is substituted to Fire module in Squeezenet network(2)Aiming at the time-consuming problem of GMM iterative modeling in the Grab Cut algorithm to solve the t?links weights,the PNN network is introduced to replace the GMM model.The PNN network has no parameter training and no iterative solutionmodel,which can greatly improve the segmentation efficiency of the Grab Cut algorithm.The hidden layer center vector of PNN network is determined by the training samples.When the training samples are selected,this paper constructs the foreground and background gray histograms and selects the pixels with higher pixel values as the training samples to improve the prediction ability of the PNN network.(3)The improved Faster R-CNN and Grab Cut algorithms are combined and applied to the water gauge detection.The algorithm can automatically process the water gauge image(or video)of reservoir,tunnel,drainage and other scenes with high universality.The water level detection system based on the fusion of Faster R-CNN and Grab Cut is designed and implemented.After testing,the system is stable,and has been applied in practical projects of cooperative units.