Research On Target Detection Technology Of High Resolution Remote Sensing Image Based On Neural Network

Images acquired by remote sensing technology cover a wide range of landmarks and features which contain a lot of valuable information.Remote sensing image interpretation is to judge the natural landforms,artificial landforms and target information within the coverage area of the image through the acquired remote sensing image data,which is widely used in civil fields such as resource exploration and environmental monitoring,as well as in military reconnaissance fields such as obtaining the dynamics of hot spots and understanding the deployment of enemy forces.With the constant improvement of remote sensing image resolution and the wide application of neural network in image processing,the technology of small target detection based on remote sensing image,the technology of target depth feature extraction based on neural network and the technology of target detection have been greatly developed.Focus imaging reconnaissance this hotspot research direction in the field of electronic countermeasures,combined with the project requirements,based on the high-resolution optical remote sensing images,samples of remote sensing image data set expansion technology,remote sensing images based on weak supervision and learning target positioning technology and remote sensing image target recognition technology based on hybrid attention mechanism study,In order to realize the remote sensing image target detection in the wide field of vision.Firstly,in view of the limited scale of remote sensing image database and the high labor cost of establishing it,the shuffle-attention GAN network is proposed based on GAN network by combining the improved self-attention module and channel Attention module with GAN network infrastructure.The network can expand remote sensing image data set based on a limited remote sensing images database.The experimental results show that compared with the existing GAN networks,the shuffle-attention GAN network generated false sample images with higher quality and higher detail.Secondly,aiming at the high cost of manual tagging required by the current database of strong supervised learning and the low precision of target detection technology of weak supervised learning neural network,this thesis puts forward that the weak supervision and target detection based on the convolution neural network,using convolution layer’s ability to locate features,using heatmap to mark convolution in the corresponding category area,using the deconvolution accuracy and to distinguish the target layer further refine area.The weakly supervised target detection network proposed in this thesis can realize the target detection and location under the image level weakly supervised label,and at the same time ensure the high efficiency and high accuracy of target detection.Finally,in view of the remote sensing image classification methods lack target accurately,on the basis of the dataset expansion,combined with the reality of military requirement,this thesis proposes a target classification of network based on hybrid attention mechanism to realize the objective of the remote sensing image classification precision and explore the influence of attention mechanism and learning rate on neural network classification.Experimental results show that,compared with other deep neural networks,the mixed attention mechanism residual neural network adopted in this thesis has higher classification accuracy and lower model complexity.