Research On Remote Sensing Image Segmentation Algorithm With Multi-Feature Fusion Under Integration Strategy

Remote sensing image segmentation plays an important role in the process of intelligent interpretation of remote sensing images.With the rapid development of artificial intelligence deep learning technology,more and more scholars have incorporated deep learning ideas into remote sensing image target segmentation tasks.However,the current algorithms are difficult to meet the needs of practical applications for the following reasons:(1)the target data is single.Due to the large imaging range of remote sensing images,it leads to a large workload and tedious steps when segmenting multiple types of feature targets,so more literature only segments and extracts one or two types of features,which cannot achieve the full coverage of remote sensing image information.(2)There are more interfering objects with similar visual characteristics such as texture and color.Unlike other natural imaging,the complex imaging process,wide field of view and rich content information of remote sensing images have increased the inter-class similarity of targets to a certain extent,leading to confusion in segmentation results.(3)The target changes with season,light,scale and other reasons.Common metrics to describe the features of remote sensing images include pixel,spectrum,length,edge,shape,shadow,semantics,etc.,and these metrics change with the shooting conditions,so the method of designing features only by experience often can only deal with specific data and cannot be truly automated.(4)Severing the connection between target foreground and background information in the image.Most segmentation networks use the feature map after the last layer of convolution for prediction,meanwhile,the background information of the image is not considered,resulting in inaccurate target semantic features and blurred segmentation boundaries.To address the above problems,this paper proposes a remote sensing image segmentation algorithm with multi-feature fusion under integration strategy based on Mask R-CNN instance segmentation network and Seg Net semantic segmentation network,and tries to propose an intelligent method to solve the problems of feature confusion,feature information loss and blurred segmentation boundary in the process of remote sensing image segmentation.The specific research contents are as follows:(1)To address the problem that different targets have different degrees of feature response,a weight-back attribute feature attention mechanism is added to the neck end of the original Mask R-CNN instance segmentation network.This mechanism can target the shape,texture,size and color features of remote sensing image feature targets,and back-propagate the weight values of feature responses to make the information circulate among networks,improve the utilization rate of features,and obtain the image foreground THINGS class segmentation map.(2)To address the loss of shallow features due to pooling operation,a densely connected semantic feature enhancement operation is added to the original Seg Net semantic segmentation network.This operation is based on the "jumping idea" to establish the dense connection between all the layers in front of the network and the layers behind.For the irregularly shaped targets,the feature enhancement module is used to fuse the global and local contextual information to improve the boundary feature expression of the target and obtain the image background stuff class segmentation map.(3)In order to output the pixel-level segmentation results of the whole image,complete the effective integration of the two segmentation maps,and improve the robustness of the model,this paper proposes a dynamic weighted integration strategy based on the voting method at the end.The strategy adopts the super-half voting method to output the maximum probability value of the target segmentation result,and verifies the accuracy of the voting result through dynamic weighted integration to realize the pixel-level segmentation of remote sensing images.(4)In order to prove the effectiveness of the remote sensing image segmentation algorithm with multi-feature fusion under the integration strategy proposed in this paper,the segmentation experiments are verified by using homemade datasets for 10 feature types,and the experimental results are evaluated in terms of accuracy from both the model itself and module comparison,and the advantages and disadvantages of the method in this paper are comprehensively analyzed.When the algorithm of this paper is initially applied in practical tasks,it can provide correct segmentation results with clear boundaries and good overall segmentation quality,which proves that the algorithm of this paper has certain generalization ability and can provide a new solution for the intelligent interpretation of remote sensing images.