Research On Remote Sensing Image Segmentation Algorithm Based On Deep Convolutional Networ

With the development of optical remote sensing technology,the acquired remote sensing images have become higher and higher in resolution,providing rich perceptual information for subsequent analysis applications.Semantic segmentation of high-resolution remote sensing images can automatically extract ground targets types and distribution information in the images,which has important application needs in urban planning,disaster warning,and environmental monitoring.Semantic segmentation based on deep convolutional networks is the mainstream method at present.Although it can bring certain improvements in segmentation accuracy,the accurate segmentation of ground targets in remote sensing images is still a difficult problem due to their anisotropic distribution characteristics such as large changes in aspect ratio and wide-scale range.In addition,the commonly used supervised learning methods require pixel-level accurate annotation,which significantly increases the cost of data annotation.To this end,this paper conducts in-depth research on anisotropic target feature extraction and sample dependency reduction,proposes a corresponding segmentation network model,and validates the performance on high-resolution remote sensing datasets,specific results are as follows:(1)A remote sensing image segmentation model based on anisotropic context fusion networks.The existing deep convolutional network models are still inadequate for segmenting anisotropic targets with long-range band structure and dense discrete distribution.This paper proposes a remote sensing image segmentation network with a fused anisotropic context.The network is designed with a boundary gradient convolution module,multi-scale parallel atrous convolution,and anisotropic composite strip pooling modules,and is integrated into an end-toend codec network,which can effectively capture anisotropic context information of targets at different scales in high-resolution remote sensing images,optimize target boundaries,and thus improve target segmentation accuracy.Experimental results on two publicly available remote sensing datasets show that the anisotropic context fusion network can achieve better performance than other advanced segmentation networks,and the ablation experiments also validate the effectiveness of the modules of the network in this paper.(2)A multi-loss semi-supervised semantic segmentation model based on contrastive learning.The supervised type of semantic segmentation network requires a large number of labeled samples for training and learning,while the pixel-level annotation of remote sensing datasets is costly.To reduce the dependence of the segmentation model on labeled data,this paper introduces pixel-level contrastive learning and proposes a multi-loss semi-supervised segmentation network model that makes full use of a small number of labeled samples and a large number of unlabelled samples to complete the semi-supervised segmentation task.Using the average teacher network as a framework,a memory bank is built through the teacher network to store high-quality feature vectors.The attention contrastive learning module is proposed based on consistency regularisation and entropy minimization,and the student network is optimized using a weighted multi-loss function.The attentional contrast loss measures the similarity between the output of the student network and the high-quality feature vectors in the memory bank,so that the output of the network is constantly similar to the same samples in the memory bank,thus improving the segmentation accuracy of the network.The experimental results show that the semi-supervised multi-loss segmentation network based on contrast learning can maintain a high segmentation accuracy at 1/30 of the labeled data.