Research On Detection Methods Of Small Targets In High-resolution Remote Sensing Images Based On Deep Learning

Optical remote sensing images with low spatial resolution can be used to detect large ships and other targets with the large physical size.However,it is difficult to use them to detect targets with small physical size because optical remote sensing images with low spatial resolution lack information of spatial details.With the continuous progress of the remote sensing technology,the spatial resolution of the remote sensing images which people can obtain becomes higher and higher.The remote sensing images with high spatial resolution have more abundant spatial texture.At the same time,the traditional representation of small objects is extended from the level of a few pixels’ description to the level of spatial description.Therefore,based on high-resolution remote sensing images,people can explore and carry out effective detection of small targets.The main work of this thesis is to study the detection method of small targets in high-resolution remote sensing images based on deep learning,which mainly includes the following contents:Firstly,starting from the basic principles of the neural network model of deep learning,this thesis introduces the relevant principles of the neural network and reveals the relationship between the convolutional neural network and image processing.Aiming at the limitations of traditional target detection methods in the extraction of candidate boxes and target classification,a multi-stage target detection network model based on deep learning represented by RCNN is introduced.Then,in view of the small pixel range of targets,this thesis explores the method of enhancing the spatial resolution of images with small targets in order to improve the precision of target detection algorithm.Based on the full research of the present methods of image super-resolution,this thesis introduces the Generative Adversarial Networks algorithm which is one of the most popular algorithms in the field of deep learning.This thesis combines that algorithm with the method of image superresolution.This work effectively achieves improving the performance of the small target detection.Finally,to make full use of the characteristic information of the small targets and solve the problem of the geometric deformation of small targets,this thesis explores the method of the full convolution neural network and combine that method with the small target detection.This thesis studies a target detection framework based on regional full convolution neural network and RCNN.In addition,a deformable convolution structure is studied to improve the fixed structure of convolution kernels,further to improve the performance of the detection of small targets.