Research On Remote Sensing Image Feature Extraction Based On Pyramid Model And Attention Mechanism

For the problem of semantic segmentation of remote sensing images,at present,the most commonly used method is to use deep learning technology,that is,to use convolutional neural network algorithms for semantic segmentation,this method has gradually replaced traditional non-deep learning algorithms.Although deep learning technology has made great achievements in the field of computer vision such as semantic segmentation,it is undeniable that the semantic segmentation technology based on convolutional neural network still has some challenges after years of rapid development,such as the difficulty of segmenting small objects.,target edge segmentation blur,etc.,especially when it is applied to remote sensing image segmentation,these problems become more prominent,so the semantic segmentation technology based on convolutional neural network has huge room for improvement.In order to effectively improve these problems,this paper is based on deep learning to explore remote sensing image semantic segmentation technology based on convolutional neural network.This article mainly does the following work.1.A U-shaped network based on the Atrous Spatial Pyramid model is proposed.The model uses DeepLabv3+ as baseline and is attached with a multi-level feature aggregation module to replace the original decoder.A number of low-level features are derived from the backbone.The low-level features from the backbone,the fusion features from the Atrous Spatial Pyramid,and the features from the upper-level feature aggregation module are first merged through the cascade structure in the decoder structure,and then they are output.The network uses multiple low-level features that contain rich spatial information which improves the ability to segment objects in remote sensing images.2.A general salient target detection model based on difference of Gaussian feature network is proposed,and applied to remote sensing image segmentation tasks.This network introduces the classic center-surround contrast theory into deep convolutional neural networks.Firstly,a difference of Gaussian pyramid structure was constructed on multiple scales of deep features to perceive the pop-out characteristics of salient targets in an image.Then,the differential feature was used to selectively enhance the informative deep features extracted from the backbone and finally the accurate segmentation of remote sensing images result was obtained.In addition,the Gaussian smoothing process was formulated as a series of standard one-dimensional convolutional layers in our network design,which enabled an end-to-end training phase while reducing the computational complexity.3.Remote sensing image usually has higher complexity,rich feature information and a large amount of redundant information,which is more convincing than public datasets with clear targets and distinct foreground and background.This paper uses a public dataset for verification.And to make the results more convincing,a remote sensing image dataset of northern planting greenhouses is proposed.The dataset takes the greenhouse as the segmentation target.The results of the experiment show that the evaluation index of the algorithm proposed in this paper is better than that of many excellent segmentation algorithms in recent years.In summary,this paper has carried out the research on the pyramid model and attention mechanism based on the convolutional neural network for the feature extraction technology in remote sensing images.A number of experiments have also verified the effectiveness of the proposed algorithm.Therefore,the research in this paper has potential application value in the related fields of remote sensing image processing.