High-resolution Image Building Target Detection Based On Information Fusion

In recent years,remote sensing technology has made great progress.With the increasing number of remote sensing images,the spatial resolution of remote sensing images continues to improve,and the spectral information is more abundant,providing important analysis conditions and resources for the research of remote sensing images in various fields.Target detection is an important part of remote sensing image processing,and its application fields cover all aspects of military and civilian use.Target detection of buildings from high-resolution remote sensing images has a wide range of applications in GIS database updates,military reconnaissance,land use analysis,urban planning,disaster assessment and other fields.Traditional building target detection mainly relies on the underlying visual features,such as color,texture,shape,etc.This type of method is slightly less universal,its detection performance depends on the low-level features of manual screening,and its expression ability is also limited.In recent years,with the continuous development of computer hardware technology and the emergence of large-scale high-resolution remote sensing image learning samples,convolutional neural networks have shown excellent detection performance in the field of target detection,greatly improving the accuracy of these fields.However,when a single feature is used for target detection,it is difficult to avoid the loss of detailed information of the building in the network learning process,which makes it difficult for the network to accurately extract and represent relevant features,resulting in unsatisfactory detection results.In this case,relevant scholars try to apply the idea of information fusion to the task of building target detection,and build a target detection network from multiple data sources,multi-scale feature fusion,and classifier decision fusion.These attempts have achieved good results.In this paper,based on the related ideas of information fusion,the following two aspects of work are carried out for the task of high-resolution image building target detection:(1)High-resolution image building target detection method based on edge feature fusionDue to the complex background of the remote sensing image and the small size of the building,it is prone to problems such as blurred and missing building outlines during the detection task.Aiming at this problem,this paper designs an adaptive weighted edge feature fusion network(VAF-Net).For the task of detecting buildings in remote sensing images,this method has carried out the following work.The first is based on the encoding and decoding network U-Net to construct the basic network UV-Net for target detection.By increasing the number of network layers and the "convolution-pooling" structural block,the network parameters are appropriately increased,and the U-Net network has too few parameters.The second is to design the construction of the building target detection network VF-Net that integrates edge features.Based on the results of the edge detection of the Sobel detection operator,combined with the basic detection network UV-Net,design parallel edge features Detect the subnet and fuse two different feature maps at the corresponding level to enrich the edge information in the feature maps;third,an adaptive weighted feature fusion strategy is designed for the feature fusion part.With the help of network backpropagation,it can be self-contained adaptive update fusion weights,so as to make better use of the contribution of different features to the network.(2)High-resolution image building target detection method based on super-pixel decision fusionThe above-mentioned method based on edge feature fusion solves the situation of edge blur to a certain extent,but it performs poorly in the detection of larger buildings.At the same time,affected by the shadow of the building,its edge image may be noisy.Therefore,this paper introduces the super-pixel segmentation algorithm,and does the following work: First,the optimization of the super-pixel segmentation function SLIC.Before the network design,this chapter first optimizes the shortcomings of the SLIC superpixel segmentation algorithm that cannot automatically select the K value.With the help of the super pixel optimization merging function and the CFSFDP algorithm,the K value is adaptively determined and the segmented region is optimized.The second is the design of the super-pixel decision-making fusion network.On the basis of superpixel image preprocessing,the improved basic target detection network UV-Net is used as the backbone network,and the RGB image and superpixel segmentation image are used as input to design the decision-making fusion network,and the fusion is carried out at the decision-making level.The final decision output is made with the aid of the Adaboost algorithm and the classification results of the basic classifier.On the basis of method(1),this method further improves the accuracy of building detection,and the detection effect is better.The methods proposed in this paper can improve the accuracy of high-score image building target detection to a certain extent.Compared with the basic detection network U-Net,the comprehensive evaluation index F1-score has been improved by nearly ten percentage points.It can also be seen from the output image that the detection result of the method in this paper is complete,the edge of the building is clear,the detection of small buildings is more effective,and it has good practical value.