Research And Application Of Semantic Segmentation Method For High Resolution Remote Sensing Images Based On Deep Learning

Semantic segmentation is a key step in image interpretation,which plays a decisive role in the accuracy of subsequent processing of remote sensing images.Due to the characteristics of high spatial resolution remote sensing images such as rich detail information and diverse feature structures,semantic segmentation becomes a difficult task in the process of remote sensing image analysis and processing.In the research,the semantic segmentation of high resolution remote sensing images is studied,and the segmentation model is applied to the development of a prototype of high resolution image road safety monitoring and early warning system,and the road extraction is completed by the method proposed in this research.The specific research work is as follows.(1)Semantic segmentation method of high resolution remote sensing images based on encoding-decoding structure.In order to improve the interference of " the same object different spectrum " and " the foreign matter same spectrum " on the semantic segmentation of high resolution remote sensing images,a semantic segmentation method based on encoding-decoding structure is proposed.Before the encoder pooling and decoder upsampling,the attention module is used to weight the information of the feature map in channel and spatial dimensions to highlight the important features while filtering out the classification noise.Secondly,an optimized pyramid pooling model is added at the bottleneck of the encoder-decoder interface,and the pooling operations at different levels are used to expand the receptive field of the model and fuse the multi-scale features to extract the global contextual information of the feature map.Finally,the dataset used for the experiments is analyzed for the characteristics of semantic segmentation labels,and the misleading effects of incorrect annotation on the model during training are reduced with the help of label smoothing loss function.(2)Research on semantic segmentation model of high resolution remote sensing images based on ensemble strategy.In order to improve the misclassification of some features by a single model and improve the semantic segmentation accuracy of the high resolution remote sensing images further,the concept of ensemble learning is introduced,and the weighted voting strategy is used to reduce the final inference variance of the model.The experimental results prove that the proposed method can effectively reduce the generalization error of the model,the segmentation result is closer to the real situation,and the misclassification of some pixels by a single model is improved.(3)Prototype design and implementation of high resolution image road abnormal state monitoring and early warning system.By optimizing the process of road safety monitoring based on computer vision method,the process of monitoring and early warning of abnormal state of high resolution image road is proposed,and a large-scale road safety monitoring system based on visible remote sensing is formed,and the corresponding system prototype is developed based on this system.In the process of system development,the previous algorithm is encapsulated into a road extraction module,and a high resolution image road extraction dataset is produced.Through the deep model training method based on transfer learning and data augmentation,the problem of time-consuming and labor-intensive pixel-level annotation of the semantic segmentation dataset is effectively solved,which provides the basic data for the subsequent detection of road pavement diseases.