Research On Multi-target Detection And Segmentation Algorithm Of Remote Sensing Images Based On Deep Learning

Remote sensing images target detection and segmentation is an important research direction of computer vision,and it has a very wide application prospect in civil and military fields.With the development of technology,the use of deep learning to process remote sensing images has become a hot spot in current research.The networks of deep learning target detection and segmentation can be divided into two categories based on whether the region of interest extraction network is used or not,namely the high-precision and low-efficiency two-stage network with the interest extraction network and the low-precision and high-efficiency one-stage network without the interest extraction network.This thesis focuses on the two-stage remote sensing images multi-target detection and segmentation algorithm.The main tasks completed are as follows:(1)Multi-target detection in remote sensing images.Firstly,in order to solve the problem of remote sensing images feature extraction,a multi-layered feature pyramid structure with layered jumping feature fusion is designed to enhance the fusion of the underlying information of the images and the high-dimensional abstract information.Secondly,in order to solve the problem that the size difference among remote sensing images is too large and there are too many small targets,it is very easy to cause insufficient feature extraction and the problem of missing detection of the area of interest of large or small targets.By adding the K-means algorithm to take advantage of the size characteristics of the remote sensing images themselves,it automatically initializes the area collection of the region of interest extraction network.It makes the network better take into account the size characteristics among different targets,and it is easier to implement subsequent target detection.Finally,for some small targets in remote sensing images,there is too much overlap among them.In order to avoid the defect that non-maximum suppression algorithms is easily used to cause target loss,a softened non-maximum suppression algorithm is used to reduce the probability of each target region being deleted by reducing confidence instead of hard deletion,and improving the accuracy of detection.(2)Multi-target segmentation in remote sensing images.Firstly,in order to improve the ability to extract the target boundary of the network,a partially fused fully connected network is added on the basis of the fully convolution network.The more accurate extraction function of the fully connected network is used to assist the fully convolution network to achieve better segmentation.Secondly,To ensure that each pixel can be classified on a feature map with higher resolution,an upsampling operation combining bilinear interpolation and 1×1 convolution is designed to replace the common deconvolution method,eliminating the checkerboard effect brought by the deconvolution network and reduces the network’s computing volume.Finally,in order to further strengthen the network’s target positioning and boundary segmentation capabilities,a self-attention mechanism module and an edge detection operator are introduced,and combines them to achieve better segment.(3)In addition,based on the NWPU VHR-10 dataset,the algorithm in this thesis is compared with other methods of the same kind and related experiments to verify the work.The experimental results show that,in terms of target detection,the algorithm in this paper significantly improves the detection accuracy.When the IOU is 0.7,the average accuracy rate reaches 90.99%,especially for small target detection accuracy.In terms of target segmentation,the average segmentation accuracy reached 95.68%,92.25% and 79.35%when the IOU was 0.5,0.6 and 0.7,and the segmentation network in this paper can extract the target edge contours in more details.