Design Of A Deep Learning-based Semantic Segmentation System For Remote Sensing Images

Semantic segmentation of remote sensing images is a very valuable computer vision task.Its main purpose is to classify each pixel of the input image.Due to the higher demand for remote sensing information in real life,the research on semantic segmentation of high-resolution remote sensing images in the field of computer vision has become increasingly active,and it is widely used in the fields of land resource monitoring,road extraction and land division.In recent years,the application of deep learning in the field of remote sensing image semantic segmentation has achieved great success,which has promoted its speed in parsing remote sensing image information and extracting substrate features,and has also greatly improved the processing of remote sensing.The accuracy of image-related tasks.However,the characteristics of complex imaging,information redundancy,and diverse types of remote sensing images make it face great challenges in the application process.In addition,the application of semantic segmentation in the field of remote sensing also faces the same problems as other application fields,such as the huge scale difference between the target objects in the input image and the loss of accuracy at the expense of low-level position information.Therefore,solving such problems has become a focus of many researchers.This article mainly focuses on the application of deep learning semantic segmentation algorithm in remote sensing images,and on this basis,designs a remote sensing image semantic segmentation system to conduct in-depth research on the application of remote sensing image semantic segmentation.The main work is as follows:(1)In the process of semantic segmentation,the increase of the channel capacity of the deep convolutional neural network will bring rich feature information,but the space capacity will be correspondingly compressed,which makes the prediction result inaccurate in the segmentation of feature details.To solve this problem,this thesis proposes a semantic segmentation network based on the regional attention mechanism,which is used to divide the target features of high-resolution remote sensing images.The regional attention mechanism network proposed in this thesis follows a semantic segmentation encoding-decoding architecture,and includes the following three strategies to improve segmentation accuracy: an enhanced GCN module is used to capture the semantic features of remote sensing images;a multi-scale grouping fusion module adopts Different sampling densities are used to capture different contextual information;the regional attention module assigns larger weights to high-value information in different regions of the feature map.This thesis verifies the proposed remote sensing image semantic segmentation network on Potsdam dataset and Jiage dataset.The experimental results show that the model results using these optimization strategies are better than the basic encoding-decoding network.Among them,the F1 score,average intersection ratio and pixel accuracy on the Potsdam dataset increased by 10.81%,19.11% and 11.36%,respectively,and increased by 29.26%,27.64% and 13.57% on the Jiage dataset.(2)In order to verify the practicability and effectiveness of the semantic segmentation network model proposed in this paper,and strengthen the technical interaction between remote sensing image processing and deep learning applications,this paper creates a remote sensing image semantic segmentation system based on flask architecture,which is sufficient Embedded deep learning algorithm,and on this basis,combined with the technical advantages of the B/S framework and database,further verify the practicality of the algorithm model proposed in this paper.By analyzing the requirements of the system,relevant modules for user management,data set management,semantic segmentation network management,and parameter model management are designed.And on this basis,two aspects of the system's data layer and API interface processing layer are implemented.Finally,a test case for each module function is designed by simulating general user needs,and a black box test method is used to test the system.The test results show that the system can meet general user needs,save the cost of study and research,and is simple to operate.,Features of low technical complexity and high modularity.