Optical Remote Sensing Image Object Detection Based On Contourlet FPD-RFCN

Optical remote sensing image is widely studied due to its simple imaging principle,diversified image acquisition channels,low acquisition cost and other advantages.At the same time,optical remote sensing image contains abundant image information and is widely used in such fields as urban construction,enemy situation monitoring,agriculture and forestry construction.Deep learning can extracte features automatically from the data effectively,and through the reasonable design of structure optimization of the performance of specific tasks and popular with researchers.Although deep learning technique is applied to object detection has been a huge breakthrough,but due to the imaging methods of optical remote sensing image data characteristics,the existing methods is usually a high miss rate,low accuracy problem.In this thesis,Deformable Convolution and Deformable Pooling,Nonsubsampled Contourlet Transform,Group Convolution,and Reinforcement Learning based on policy gradient are applied to optical remote sensing image object detection tasks to improve the accuracy and reliability of optical remote sensing image object detection model,thus promoting the development of optical remote sensing image object detection technology.The main research content of this thesis is as follows:1.Aiming at the problem of small object detection difficulty and large target deformation difference,a object detection algorithm based on feature pyramid networks structure for deformable convolutional networks(FPD-RFCN)is proposed.This method using the networks by means of characteristics of the pyramid of multilayer characteristics,improve the performance of the small target detection,the characteristics of different stages for deformable,convolution,use can extract the characteristics of deformation of convolution deformation can be automatically learning objectives,so this design method through the networks structure of effective solves the two problems of optical remote sensing image application,and only a few of the amount of calculation and parameter is introduced.Four different datasets from the Quick Bird and Jilin 1th satellites were used to carry out experiments respectively,and compared with other latest target detection methods.Experimental results show that this method is effective in solving the problem of remote sensing image target detection from the perspective of networks structure design,this research was used to obtain the excellent award in the remote sensing image competition of National Natural Science Foundation of China and the special award in the IPIU object detection competition.At present,the algorithm needs to further improve the design of networks structure and simplify the networks structure,so as to improve the accuracy and optimize the detection speed at the same time.2.For multi-scale transform,Nonsubsampled Contourlet Transform(NSCT)can effectively enhance the edge features,and at the same time,in order to prevent the transformation from losing the texture information in the original image,a grouping FPD-RFCN object detection method based on NSCT is proposed.In this method,NSCT is firstly used for edge enhancement,and channel deconvolution of the enhanced data and the original data is carried out to serve as the input of FPD-RFCN.This model not only makes effective use of multi-scale transform,but also puts forward a grouping convolution mode of channel dispersion,which makes reasonable and effective use of the different advantages of images before and after transformation.Finally,experiments are carried out with data from different satellites,and detailed comparative experiments are carried out with other methods.The experimental results prove that the edge features extracted by using group convolution in this scheme are conducive to improving the detection accuracy,At the same time,the research is applied for the national invention patent.The algorithm needs to further optimize the NSCT to obtain a transformation that is more conducive to extracting image features.3.When there is a deviation between the distribution of training data and test data,the FPDRFCN object detection method based on reinforcement learning is proposed.This method mainly USES reinforcement learning to be good at learning from the environment to optimize long-term goals.Reinforcement learning is applied to learns data representation and data matching jointly from image source domain and target domain,at the same time using depth convolution neural networks learning algorithm as the main body of object detection,through iterative optimization,eventually converge to a better effect.Finally,experiments are carried out in four groups of data and different methods.The experimental results show that the algorithm can effectively optimize the training of the model and reduce the risk of model overfitting training data.Meanwhile,we presented the relevant theories and other experiments of the Adaptation method of the reinforcement Learning Domain to an article on the International Conference of Learning representation-2018(Domain Adaptation via Distribution and Representation Deformation: Training Data Selection by Reinforcement learning?International Conference of Learning Representation).At present,the problem of this algorithm is how to optimize the training speed of the model.