Remote Sensing Image Segmentation Algorithm Based On Adaptive Feature Fusion Network

Remote sensing image segmentation is widely used in many fields,such as natural resource management,disaster detection and urban planning.For the task of remote sensing image segmentation,most of the existing convolutional neural networks focus on mining the deep semantic features of complex objects in remote sensing images.However,this ignores the shallow high-resolution features including a large number of small targets and edges,which to some extent affects the segmentation effect of existing networks on remote sensing images.To better fuse shallow high-resolution features with deep semantic features and improve the segmentation effect of remote sensing images,the main research of this paper is as follows:(1)A remote sensing image segmentation model based on encoder-decoder structure is proposed.Firstly,a multi-scale feature extraction network is built by stacking multiple residual combination modules in the encoder to extract shallow highresolution features and deep semantic features.Through the decoder composed of several de-convolution modules,the deep semantic features are gradually resolved to the original high-resolution feature map.To make full use of the shallow highresolution features,the deep features are fused with the corresponding shallow highresolution features through the design of long-skip connections,so that the shallow high-resolution feature can be directly conveyed to the prediction results.Experimental results and comparative analysis show that the proposed model is effective in remote sensing image segmentation.(2)To make full use of high-resolution features,on the basis of(1),this paper proposes an adaptive feature fusion network based on pixel encoding from the perspective of improving the fusion efficiency of spatial level features.It creatively introduces self attention mechanism to optimize the fusion of shallow and deep features in image segmentation network.However,directly using the original self attention in feature fusion will bring high computational complexity,so its structure is reasonably modified in this paper.A pixel encoding structure is designed,which optimizes the global dependence to vertical and horizontal information dependence in the structure.Then,through the second iteration of pixel encoding structure,the interaction of global arbitrary pixel information can be realized again,which realizes adaptive feature adjustment.Multiple ablation and contrast experiments show that the adaptive feature fusion network based on pixel encoding can effectively complete remote sensing images segmentation.(3)To improve the segmentation accuracy of remote sensing images and select more representative information from the shallow and deep fusion features,on the basis of(1),this paper proposes an adaptive feature fusion network based on channel encoding from the perspective of channel level feature fusion.This is the first time to incorporate channel attention into the cross-level feature fusion.To make the original channel attention module more suitable for the remote sensing image segmentation task,a new channel encoding structure instead of channel attention structure is embedded in the feature fusion.Firstly,to improve the ability of distinguishing shallow features and deep semantic features,global maximum pooling is used instead of global average pooling in the squeeze module,which enhances the feature extraction of the difference.Secondly,to maintain the one-to-one correspondence between channel weights and feature dimensions,and avoid calculation consumption caused by all-channel information interaction,the two full-connection layers are repalced by one-dimensional convolution of channel dimensions.In this way,the information interaction required by the network can be effectively completed and the segmentation effect with high precision can be achieved.The experimental results show that the adaptive feature fusion network based on channel encoding can accurately and efficiently complete the remote sensing image segmentation task.