Remote Sensing Image Classification Research Based On Deep Learning

Object detection and semantic segmentation of high-revolution remote-sensing image are two hot research fields of remote-sensing image classification technology.In recent years,deep learning algorithm has achieved great success in the field of computer vision.Both in object detection and semantic segmentation tasks,it has surpassed the algorithms of traditional shallow machine learning.However,remote-sensing images are different from ordinary streetscape images: In terms of semantic segmentation,remote-sensing images have the characteristics of complex imaging and diversified information;in the terms of object detection,remote-sensing images have the characteristics of differences in target size and dense distribution.If the existing algorithms are directly applied to remote sensing-images,the results obtained are not ideal.Based on the characteristics of high-resolution remote-sensing images,this paper designs and improves two different networks for semantic segmentation and target detection.The specific content is arranged as follows:(1)A semantic segmentation network of feature-fusion remote-sensing image based on down-sampling is designed.The network extracts low-level semantic features by introducing the high-resolution original image into the down-sampling module,and based on these low-level semantic features,further extracts details of multiscale high-level semantic features through Mobilenet V2 network and spatial pyramid pooling layer.Finally,the low-level semantic features and high-level semantic features are fused to reduce the network complexity,avoid the occurrence of over-fitting phenomenon,and reduce the calculation amount.Through the training and testing on the open data set,the network surpasses the other four classical semantic segmentation networks in terms of the global accuracy.(2)Being aimed at its problem of inaccurate localization of small targets,the classic Faster-RCNN network is improved to be more suitablefor small image detection.There are two specific improvement measures:firstly,our network add a positioning refinement network after the original Faster-RCNN network.The refinenet continues input the feature maps obtained from the convolutional neural network and the results obtained by the regression layer into the Ro I Pooling layer to refine the positioning and thus,improve the positioning accuracy;secondly,The candidate frames generated in the RPN network were adjusted from 9 to12 with different sizes,making the improved network more suitable for small object detection in remote sensing images.The average accuracy rate of the improved network is 4% higher than that of Faster-RCNN network.For smaller objects,such as vehicles,the average accuracy is improved by 6.8%.