Research On Scene Classification Of High-Resolution Remote Sensing Image Based On Saliency Joint Convolutional Neural Network

Scene classification of high-resolution remote sensing image is an important work of interpreting remote sensing image information.In order to identify the information contained in remote sensing images more comprehensively and accurately,researchers are not satisfied with the traditional artificial design features.They apply deep learning to scene classification and construct convolutional neural network to detect and classify remote sensing images.At present,scene classification based on deep learning has become a mainstream method,which is widely used in the fields of computer vision such as semantic segmentation,object detection and image classification.However,high-resolution remote sensing images often have the characteristics of strong background complexity,large diversity of targets and difficulty in distinguishing between target information and background information,which brings many difficulties to scene classification tasks 。In order to extract the target information better,make full use of the salient features and deep semantic features of the image to improve the classification accuracy,a novel scene classification method of high resolution remote sensing image based on saliency features and convolution neural network is proposed.Firstly,saliency features are extracted from the original image based on saliency detection to obtain saliency map;then,the target region and background region are divided by binary processing,and the deep semantic features are obtained by extracting the region of interest of the original image;finally,the images containing the region of interest are constructed as sample pairs A deep full convolution neural network model is trained to obtain deep-seated features,and softmax is used to classify the deep-seated features,and the detection model is used to classify the sample data.Experiments on uc-merced21 dataset and whu-rs19 dataset show that the proposed model can effectively distinguish the main target information from the background information,reduce the interference of irrelevant information,reduce data redundancy,and effectively improve the classification accuracy compared with the existing methods.This method has the characteristics of high classification accuracy,fast running speed and strong robustness,which can be applied to scene classification tasks of high resolution remote sensing images.This paper has 31 figures,9 tables and 53 references.