Remote Sensing Image Segmentation Algorithm For Hidden Gas Danger Target Recognition

Buildings,structures,and construction sites near gas pipelines can cause hidden dangers.Manual inspection is one of the main ways to eliminate hidden gas dangers.In recent years,with the development of hardware,the processing speed of GPUs in the field of matrix operations has improved significantly,and more and more data is available in the era of big data.Driven by these developments,deep learning has been developing rapidly after breaking through the theoretical bottlenecks,and researches on computer vision based on deep convolutional neural networks become more and more,especially the related research based on remote sensing images.Image segmentation based on deep CNN is used to identify buildings and structures in remote sensing images to determine the location of hidden dangers with pipeline information,which can effectively improve the inspection efficiency and detection accuracy of hidden gas dangers.Therefore,the main work of this paper is to study the image segmentation of hidden gas danger related objects in remote sensing images,focusing on the building category.This thesis studied semantic segmentation and instance segmentation of building category in remote sensing images.This thesis proposed a semantic segmentation network with better performance on remote sensing image dataset based on DeepLabv3+.On the basis of the semantic segmentation model,this thesis referenced the main idea of TernausNetV2 and made some improvements,and reused the updated semantic segmentation neural network as the backbone of the instance segmentation network.At last,this thesis proved that the upgraded instance segmentation network obtained better accuracy on the SpaceNet remote sensing image dataset compared with TernausNetV2.The main work of this paper is listed as follows:1.Based on the semantic segmentation model of DeepLabv3+,this thesis improved the accuracy of semantic segmentation of buildings on remote sensing images.By adding more skip connections between the encoder and the decoder,the decoder of the network can make full use of the context information from the encoder,so that the high-level semantic features of the decoder and more low-level boundaries on the encoder can be concatenated.More location information finally made the semantic segmentation network obtain a more accurate segmentation boundary.The value of the intersection over union index of the new model proposed in this thesis is about 1.39%higher than that in the DeepLabv3+model on the SpaceNet dataset.2.This thesis reused the idea of TernausNetV2,that is,after obtaining the result of semantic segmentation,the network can directly obtain the result of instance segmentation through calculations.This thesis integrated these calculations into the neural network so that the entire process of instance segmentation can participate in the iteration of the neural network.On the SpaceNet dataset,the new instance segmentation outperforms the effect of the instance segmentation algorithm in TernautNetV2,and the F1 score is improved by 1.31%.Then,in order to further improve the performance,the watershed transform based on the semantic segmentation result of the building was changed to depend on the semantic segmentation and the multi-scale morphological gradient of the original remote sensing image.Finally,the F1 score was improved by 1.49%.This thesis performed experiments on the SpaceNet remote sensing image dataset,and a better semantic segmentation network is obtained compared with the original DeepLabv3+,and the backbone structure of the proposed better new semantic segmentation network are used.After the modification,the new instance segmentation network has better segmentation performance,and it has a 3.65%improvement in F1 score over the original TernausNetV2 algorithm.