Research On Semantic Segmentation Of Remote Sensing Image Based On Encoder-decoder Network

At present,with the update and development of satellite and aviation equipment,the quantity and quality of remote sensing images are also increasing.How to use these remote sensing images to study agriculture,military and urban development is a very important issue at this stage.In this paper,remote sensing image is employed as the principal component of research,and in conjunction with image segmentation algorithm on the basis of deep learning High-resolution remote sensing data set is employed to train and test the suggested semantic segmentation network model.The main work is as follows:For the segmentation task of road surface extraction from remote sensing images,road width is an important issue,so we focus more on the boundary detail segmentation of roads.In chapter three,design a multi-scale environment awareness network based on boundary loss for pavement extraction in graffitt-based remote sensing image data sets.In this chapter,a new feature coding network is designed,which uses residual network as the backbone network.The Atrous Spatial Pyramid Pooling(ASPP)module is added to each layer of the residual network to capture the context information of the image at different expansion rates,so as to obtain the multi-scale features of the image and establish a more flexible information flow.At the same time,CABlock(Context Aggregation Block)modules were added to each layer of the backbone network in the way of interlayer residual connection to further enhance the extracted image features.In this chapter,a new loss function BF1 is used to extract the road boundary contour more effectively.Since the supervised semantic segmentation method relies on a large number of pixel-level marked images,the supervised semantic segmentation method of remote sensing images has a high cost.In chapter four,we adopt a weakly supervised segmentation method,that is,comparative learning method,which for pre-training,employs an enormous number of unlabeled samples and,for later tasks,a fewer amount of labeled samples.Inspired by this,propose a dense multi-scale feature contrast learning network(DMF-CLNet),which is divided into two parts: global feature extraction and local feature extraction.First,in the global part,DenseASPP can capture more remote sensing image context information in a dense manner than traditional ASPP without adding parameters.Secondly,coordinated attention(CA)modules are introduced in the global part and the local part respectively to improve the overall performance of the segmentation model.Then,the perceptual losses are calculated in the global and local parts to enhanced the details of the feature.In order to solve the problem that low resolution segmentation method is not suitable for high resolution image,in chapter five proposes a cascade refinement attention network(CRANet)A detailed attention module is designed,which extracts multi-scale feature maps using the network design of the encoder and decoder.To acquire more comprehensive information,the jump join section blends high-level semantic segmentation information with low-level detail information.As for the proposed cascade refinement attention network and segmentation graph of original image,the stacked cascade design method can refine the structure from rough to fine,and obtain better segmentation image.Finally,the high resolution remote sensing image data set is applied to the network,and the proposed network is used to realize the semantic segmentation of remote sensing image scene analysis.The experimental results show that the network improves the segmentation accuracy of remote sensing image.