Research On Real-time Semantic Segmentation For Remote Sensing Images Based On Light Weight Network

The support of deep learning technology is indispensable behind the semantic segmentation technology.However,deep learning requires high computing power and storage capacity of the operating platform.In many scenarios,we need to apply deep learning in mobile terminals or embedded platforms.How to apply semantic segmentation technology on devices with limited computing power,reduce the amount of calculation,and increase the running speed has become a hot topic of current research.Based on the remote sensing image semantic segmentation data set,the thesis uses deep convolutional neural network to explore and study the lightweight real-time semantic segmentation model algorithm from the perspective of convolution solution and lightweight visual attention mechanism.The research content is as follows:First,construct a remote sensing image semantic segmentation data set.The data set has 5 segmentation categories,including background categories,and uses self-built labeling software for labeling and visualization of labeled images.This data set will be used as training for subsequent model experiments.Sets and validation sets provide a data basis for subsequent experiments.Second,from the perspective of volume integration solution,a typical Deep Labv3+ semantic segmentation network is optimized for lightweight design.By lightening and improving each component of the encoder and decoder structure,a method is proposed in the encoder.A feature extraction network composed of non-bottleneck residual blocks using channel separation and rearrangement,Lite-R-ASPP is used in the decoder,and a classifier that uses deep separable convolution.The segmentation model achieved a prediction speed of 33 FPS and an average cross-to-match ratio of 68.89%.Finally,the validity and robustness of the model are verified using the VOC2012 data set.Third,from the perspective of visual attention mechanism,the lightweight fast semantic segmentation network Fast-SCNN is improved,and the hybrid domain attention mechanism,the cross attention mechanism and the maximum expectation are added to its feature extraction module.Improve the attention mechanism,use the attention mechanism to obtain global feature information,and guide the segmentation model to accurately locate.Experiments show that the fast semantic segmentation network using the cross attention mechanism can achieve a balance of speed and accuracy on the remote sensing image semantic segmentation data set constructed in the thesis,and has generalization ability.On the VOC2012 data set,the accuracy of model segmentation is higher than that of the original network,indicating that the model has a certain generalization ability.In summary,the lightweight semantic segmentation network model based on volume integral solution and the semantic segmentation network model based on visual attention mechanism proposed in the thesis can effectively perform real-time segmentation in the scene of remote sensing image interpretation.Under the scene,it still has high efficiency.