Research On Key Technologies Of Remote Sensing Image Semantic Segmentation Based On Deep Learning

With the rapid development of remote sensing technology,high-resolution remote sensing images are more easily acquired and widely used.High-resolution remote sensing images have been widely used in urban planning,environmental protection,disaster assessment and prediction,transportation navigation,military security and other fields.In these application scenarios,it is generally necessary to use computer vision technology to analyze high-resolution remote sensing images to extract effective geometric and structural information.As an important image content analysis method,semantic segmentation has always been an important research direction in the field of remote sensing image processing.Due to the characteristics of complex backgrounds,diverse structures,and rich details in high-resolution remote sensing images,traditional semantic segmentation algorithms do not perform well when applied to the processing of remote sensing images.In recent years,with the great success of deep learning,especially Deep Convolutional Neural Network(DCNN)in the field of natural image processing,it has been extended to the analysis and processing of remote sensing images,which has promoted the development of high-resolution remote sensing images related tasks,including semantic segmentation.In this dissertation,we focus on the application of DCNN based semantic segmentation algorithms in high-resolution remote sensing images.The main research contents are as follows:(1)Due to the inherent characteristics of the DCNN,when it is applied to the semantic segmentation of high-resolution remote sensing images,it generally has the problems of inaccurate segmentation on the semantic boundary and inconsistent label prediction of the sub-regions within the objects.To tackle this challenge,this dissertation proposes an algorithm architecture combining DCNN and superpixel algorithm for post-processing the segmentation results of deep semantic segmentation models.Firstly,a deep semantic segmentation model is used to predict the label of each pixel of the input high-resolution remote sensing image.At the same time,the input image is divided into different superpixel sub-regions using the quick shift superpixel segmentation algorithm,and then the outputs of semantic segmentation model and quick shift algorithm are fed into the the region voting algorithm proposed in this dissertation to further improve the segmentation accuracy.The experimental results show that our proposed post-processing method can effectively improve the accuracy of the segmentation results of the deep semantic segmentation model,especially the accuracy of the boundary segmentation and the consistency of the sub-region prediction within the objects.(2)In this dissertation,a lightweight and dual-path semantic segmentation model based on boundary supervision is proposed to solve the problems of intra-class heterogeneity and inter-class homogeneity in high-resolution remote sensing images,while improving the accuracy and processing speed of high-resolution remote sensing image semantic segmentation.The strategy of combining the multi-level,multiple spatial receptive field and global context features to encode the local and global information,is used to address the intra-class heterogeneity challenge.For inter-class homogeneity problem,the semantic boundary predicted by the boundary detection sub-network is used as an auxiliary loss to deeply supervise the training process of the network.The lightweight MobileNetV2 for image classification is modified and applied as the feature extractor to improve the speed of our proposed network.Experimental results demonstrate that the proposed lightweight semantic segmentation network has high performance,and relatively balanced inference speed.(3)Due to the property of complex background and dense targets in high-resolution remote sensing images,the cost of semantic labeling is very high.To tackle this problem,a semi-supervised semantic segmentation model based on residual structure is proposed to alleviate the dependencies of DCNN on large-scale annotated dataset.Firstly,a semantic segmentation model is pre-trained by a small number of samples with annotated strong labels,then a large number of samples with weak labels are predicted by the model,and finally a new semantic segmentation model is trained by combining samples with strong and weak labels.In this dissertation,a new semantic segmentation network based on semi-supervised learning and residual branching structure is designed,which is different from the commonly used semantic segmentation model for semi-supervised learning.The proposed network can further exploit the correct semantic information hidden in the samples with weak labels,and learn more effective semantic features,thus achieves higher segmentation accuracy.The experimental results on the ISPRS2 D semantic labeling dataset and the WHU aerial building dataset verify the effectiveness of the semi-supervised semantic segmentation network based on residual structure proposed in this dissertation.(4)The lightweight and dual-path semantic segmentation model based on boundary supervision is implemented on the embedded GPU computing module to verify its performance on embedded platform.A remote sensing target segmentation dataset is proposed,including categories of airplane,ship,and car.The algorithms proposed in this dissertation are verified on the dataset,which achieve good segmentation performance.