Semantic Segmentation Of Remote Sensing Images Based On Multi-scale Fusion And Information Enhancement

Semantic segmentation of remote sensing images is an important method in remote sensing image processing,which provides the necessary foundation for subsequent image recognition and scene understanding.The high resolution of remote sensing images with complex scenes,large differences in target size and the presence of confusing objects all bring more challenges to the task of segmentation.To address these problems mentioned above,the main research of this thesis is as follows.(1)A multi-scale adaptive weight-based information flow fusion algorithm is proposed for the semantic segmentation of remote sensing images.In order to solve the problems of target diversity and large size differences of remote sensing images.A multi-scale coding module is proposed to extract features of 5 different dimensions using 5 different sizes of atrous convolution and conventional convolution in parallel.The features of 5 different scales are combined by channel and added with their own features to obtain features of each scale as far as possible,thus alleviating the problem of large target size differences in remote sensing images.In the decoding stage,a decoding module for weight-based information flow fusion is proposed.By fusing layer-by-layer and cross-layer connections according to weights,the information flow at higher levels can be fully utilized and the redundancy of information flow caused by too many features fusion can be avoided,thus making the segmentation results more accurate.Experimental results show that MWFNet outperforms other algorithms in segmentation.(2)A multi-scale adaptive feature fusion algorithm is proposed for semantic segmentation of remote sensing images.In order to obtain multi-scale contextual information and efficiently fuse high-level and low-level semantic information to solve the problem of excessive differences in target sizes in remote sensing images.The multi-scale context extraction module employs two layers of atrous convolution with five different dilation parameters and global average pooling in parallel to generate different scales of context information and combine the multi-scale context information by channel.In addition,to enhance the effect of feature fusion,the channel attention is embedded into the semantic feature fusion at both low and high levels.Firstly,the low-level and high-level features are combined and the global features of each channel are obtained via global averaging pooling.Secondly,the obtained global features are used as channel weights to obtain the weight information of each channel through adaptive learning at the fully connected layer.Then,the obtained weights of each channel are applied to the fused features.The features of each channel are adjusted according to the weights to achieve efficient fusion.The experiments show that the MANet outperforms other algorithms in terms of segmentation performance.(3)An adaptive information enhancement algorithm based on HRNet is proposed for semantic segmentation of remote sensing images.The algorithm includes an adaptive spatial information enhancement module and an adaptive channel information enhancement module.The adaptive spatial information enhancement module performs matrix multiplication with the feature map through the transpose of the feature map,and then uses softmax to obtain the normalized information-weighted map of the features.The matrix multiplication of the information weighted map transpose of the features with the feature map enables the dependency of global information to be captured at any position in the feature map.Finally,this feature weighting is added to the weighted original features to obtain the final output.The adaptive channel information enhancement module makes the features focus more on some useful channel information and choose to ignore some useless channel features.Experiments show that the AIENet algorithm segmentation outperforms other algorithms.