Research On Remote Sensing Semantic Alignment Algorithm Based On Convolution Neural Network

Remote sensing images are earth surface images obtained through aerial or satellite remote sensing techniques.In various visual applications,image alignment is one of the basic tasks,which aims to identify the same or similar content in two or more images.Remote sensing image semantic alignment technology is a technique that can align remote sensing images of different times and resolutions.It is often applied in geological exploration,natural disasters,and urban planning,and plays a crucial role in improving the accuracy and comparability of geological information.Therefore,further development and research of remote sensing image semantic alignment technology is particularly important.This thesis proposes two algorithms for remote sensing image semantic alignment: one based on improved attention-based neighborhood consistency network,and the other based on saliency mechanism.1.This thesis proposes an end-to-end semantic alignment framework for remote sensing images based on an improved attention-based neighborhood consistency network.Traditional remote sensing image semantic alignment algorithms suffer from problems such as low attention to local feature information,insufficient extraction of important channels and spatial information,and ambiguous matching.To address these issues,the proposed algorithm uses asymmetric convolutional blocks and an improved CBAM attention mechanism to quickly focus on important regional features during the feature extraction stage,thereby improving the network’s discriminability and ability to locate image features.The extracted features are then fed into the matching stage,where the Pearson correlation algorithm is used to establish a correspondence between the two images.In addition,a multilayer perceptron(MLP)structure is introduced into the neighborhood consistency network to establish strong matching locality,allowing information to aggregate at different positions and channels,reducing alignment bias,and enhancing matching robustness.Finally,affine transformation is performed using transformation parameters to obtain the final semantic alignment result.2.This thesis proposes a remote sensing semantic alignment algorithm based on saliency mechanism.In the feature extraction stage,the channel attention is used to improve the network in the network,identifying important channels and learning more complex feature information.Subsequently,a multi-scale bidirectional consistency algorithm is designed to supplement different matching information and establish the relationship between the source image and the target image.At the same time,the loss function is improved by using the four parameters obtained by parameter regression design to iteratively optimize the network model,in order to output more accurate semantic alignment results.To verify the effectiveness of the algorithm,experiments are conducted on a remote sensing image dataset.The experimental results indicate that the end-to-end algorithm proposed in this thesis can improve the accuracy and robustness of semantic alignment in remote sensing images compared to traditional methods and deep learning-based methods proposed in recent years.In addition,this thesis discusses the applicability of the algorithm under different circumstances and proposes some future research directions.