Research On Multi-source Remote Sensing Image Target Detection Based On Deep Convolutional Neural Network

With the rapid development of remote sensing technology,scene applications based on remote sensing images have been reflected in all walks of life.In recent years,theories and algorithms in the field of artificial intelligence(AI)have made phased breakthroughs,which has made the integration of remote sensing science and technology and computer and other interdisciplinary developments increasingly close.In actual scene applications,making full use of the characteristics of synthetic aperture radar images and optical images to perform information extraction,scene classification and recognition,specific target detection and semantic segmentation on multi-source remote sensing images has very important research value.Under the premise of continuous improvement of neural network theory,the proposal of backpropagation algorithm makes it possible to continuously extend the width and depth of neural networks,especially deep convolutional neural networks(Deep Convolutional Neural Network,DCNN)with neural network autonomous learning features The idea of replacing the previous tedious and complicated feature engineering,thus realizing the real data-driven feature engineering.In this paper,DCNN is used as the basis of the target detection framework and is organically combined with multi-source remote sensing image target detection tasks.In view of the scarcity of remote sensing data sets,a small-scale remote sensing data set was established;due to the imaging characteristics of synthetic aperture radar images and the large gap between image information expression and optical imagery,this paper carried out the reconstruction of the data set pre-selection box clustering algorithm;in addition;Richer texture information is also a major feature of multi-source remote sensing image data.In order to better mine the texture information and texture features in the image,the Gabor filter is used to reconstruct the convolution kernel;at the same time,in order to improve the multi-source image data.The problem of missing detection of small targets at the edge of source remote sensing images is improved.The loss function is improved,and edge control factors are added to the position loss of the target,the confidence loss of the target and the classification loss.Finally,due to the large size and small scale of the remote sensing image It is determined that the small objects in the image are mostly small.In order to improve the detection effect of small objects,an upsampling multi-scale detection strategy is introduced.The later-level feature maps are upsampled and merged with the previous-level feature maps of the same size.This combines both high-level semantic features and It can be well optimized for the difficult problem of small target detection.