Research On Multiscale Remote Sensing Image Object Detection Approaches Based On Cross Stage Partial Connection

With the rapid development of remote sensing technology and the successful launch of a large number of artificial satellites,a large amount of high resolution optical remote sensing images have been obtained,which contains lots of valuable information.Remote sensing image object detection technology is one of the effective means to interpret this information.In the early days,people used traditional methods for detection,but the method had a complicated process,high time complexity and poor robustness.Furthermore,it was difficult to achieve ideal results in object detection of high-resolution remote sensing images.However,the object detection method based on convolutional neural network can extract richer image features without much prior knowledge,which is the mainstream research direction of object detection.Different from natural images,remote sensing images have the characteristics of high resolution,inconsistent object scale,and complex background.More and more practical application scenarios not only put forward high requirements on the accuracy of remote sensing image object detection,but also hope for faster detection speed.Therefore,how to quickly and accurately detect multiscale objects in remote sensing images has always been the focus and difficulty of research.In order to solve the above problems,based on the existing research,this thesis focused on the accuracy and speed of multiscale object detection in high-resolution remote sensing images,and achieved good detection results.The main works of this thesis are as follows:(1)Aiming at the problem of poor detection of multiscale objects in remote sensing images,a Cross Stage Partial Multiscale Densely Connected Convolutional Networks CSPMDense Net is designed from the perspective of feature extraction and fusion of convolutional neural networks.The network combines the characteristics of a cross stage partial connection structure,a densely connected network and a multiscale feature fusion network.It has the advantages of fewer parameters,stronger feature extraction and fusion capabilities.On the basis of this network,we propose an end-to-end remote sensing image object detection method based on CSPM-Dense Net.CIo U loss is used as the bounding box regression loss function to further improve the detection accuracy.We use the improved kmeans++clustering algorithm to cluster the Ground Truth to obtain a priori anchor boxes that are closer to the true sample distribution.Aiming at the problem of the small number of public remote sensing images and the imbalanced distribution of object categories,different data enhancement and expansion methods are adopted to make the network model obtained by training have good generalization.The experimental results show that the method in this thesis achieves a m AP value of 96.78% on the NWPU VHR-10 data set and 94.33% on the RSOD data set.For 512×512input images,the detection speed is 40.40 FPS and 35.57 FPS,respectively.Compared with other methods,the method in this thesis has obvious advantages in detection accuracy,which can effectively improve the detection effect of multiscale objects and achieve the speed of real-time detection.(2)Aiming at the problem of unbalanced detection accuracy and speed in remote sensing image object detection methods,a lightweight network model CSPM-OSANet is designed.The network includes three parts: the backbone network CSP-OSANet,the parallel feature pyramid structure and the detection sub-network.The backbone network is composed of a spatial pyramid pooling module and three OSA modules(CSP-OSA)which combine cross stage partial connection structures.The CSP-OSA module can extract richer features with fewer layers,reduce the cost of network memory access,and have higher GPU-Computation Efficiency.The parallel feature pyramid structure includes three parallel branches.Each branch,which is responsible for detecting objects of various scales,adopts different dilation rates of dilation convolution to obtain different receptive fields.On the basis of this network,this thesis proposes an end-to-end remote sensing image object detection method based on CSPM-OSANet.In addition,to realize the reliable output of the bounding box,improve the positioning accuracy and improve the detection accuracy,Gaussian modeling is performed on the output of the bounding box in this method.The experimental results show that the method based on CSPM-OSANet is compared with the method proposed in Chapter 3,the m AP value on the NWPU VHR-10 data set is reduced by 1.77%,but the detection speed is increased by 309.90%.On the RSOD data set,m AP decreased by 2%,and the detection speed increased by 367.16%.Compared with other comparative experiments,the method in this thesis achieves a good balance between detection accuracy and speed.Finally,Jetson AGX Xavier on embedded platform is used to verify the portability of the method.The results show that the method we presented have higher detection accuracy on the embedded platform,and can basically meet the requirements of real-time detection.