| In recent years,the problem of urban stagnant water has deteriorated dramatically.Under the background of smart city,timely detection and early warning of the stagnant water area play an important role in ensuring the normal operation of the city.The existing detection methods of urban stagnant water are realized by setting ponding points and installing hardware devices.However,these methods have the disadvantages of high installation cost,small coverage and inability to detect the small area of stagnant water in practical application.In this paper,urban stagnant water is detected by the semantic segmentation technology based on deep learning,which can effectively make up for the shortcomings of the existing methods and has certain value in practical application.(1)Research on semantic segmentation algorithm based on attention and saliencyThe characteristics of urban stagnant water in the actual scene are complex in background,irregular in shape and diverse in ripples.Those influencing factors will directly reduce the accuracy of semantic segmentation,resulting in misjudgment and missed judgment of the waterlogging area.Therefore,this paper proposes a semantic segmentation algorithm based on attention and saliency(WSNet),which mainly designs the Spatial Pyramid Attention Module(SPAM)and Global Significance Module(GSM)to enhance the understanding of semantic information,so as to improve the segmentation effect of urban stagnant water area.On the one hand,SPAM enhances the understanding of image context information and increases attention to the details of stagnant water through the attention mechanism to improve the algorithm’s ability to predict the category of the segmented objects.On the other hand,GSM further enhances the context information by increasing the saliency of the characteristics of the waterlogging area to reduce the interference of external factors such as the sky and the mirror reflection of the smooth road.This paper verifies the effectiveness of the algorithm from the evaluation indicators such as visual effect and Mean Intersection over Union(MIo U).The experimental results show that the MIo U of WSNet algorithm on the actual urban stagnant water dataset(W2020)can reach 86.19%,which has higher segmentation accuracy and robustness than the original Pyramid Pooling Module(PPM).(2)Research on semantic segmentation algorithm based on super-resolution reconstructionThe resolution of the feature map in the semantic segmentation algorithm will gradually decrease with the deepening of the network layer,and it needs to be restored to the size of the original input image when performing algorithm inference.The current research work generally uses bilinear interpolation,transposed convolution or other common methods to up-sample the low-resolution feature map.These methods of directly expanding the resolution of the feature map are easy to cause the loss of detail information,resulting in inaccurate segmentation of the target object edge,and introduce irrelevant information,increasing the amount of calculation.Aiming at the characteristics of pixel level classification of semantic segmentation tasks,this paper proposes a semantic segmentation algorithm based on super-resolution reconstruction(WSPNet).It mainly introduces the super resolution reconstruction module(DS)designed in this paper into WSNet to reconstruct the high-resolution feature map with high quality,so as to make up for the loss of accuracy caused by the loss of image features.In this paper,multiple sets of comparative experiments are performed on the actual urban stagnant water dataset(W2020)to verify the effectiveness of the algorithm from the Mean Intersection over Union(MIo U)and visual effect.The experimental results show that the MIo U of WSPNet algorithm can reach 87.03%,which can further improve the accuracy of segmentation of urban stagnant water. |
