Feature Fusion Deep Convolutional Neural Network For Remote Sensing Image Object Detection And Terrain Classification

With the development of remote sensing technology,the acquisition of remote sensing images is more and more convenient.The object detection and terrain classification in remote sensing images are getting more and more attention.In recent years,the processing and interpretation in remote sensing images has been further developed.Improving the performance of remote sensing images object detection and terrain classification is a very significant research topic.Aiming at the problems existing in the existing remote sensing image object detection and terrain classification algorithms,this thesis proposes two new remote sensing image object detection algorithms and a new remote sensing image terrain classification algorithm based on convolutional network feature fusion.The main tasks are as follows:1.Aimed at feature maps with different sizes depending on the depth level in the convolutional network,a remote sensing image object detection algorithm based on bilinear interpolation Feature fusion Single-Shot Refinement neural network Detector(F-SSRD)is proposed.This algorithm is mainly for the shortcomings of the common object detection algorithms for small object detection.The algorithm constructs a feature learning and fusion network.The network is divided into feature fusion module and detection module.Compared with the traditional target detection method,the key of this method is to improve the resolution of deep small-scale feature maps by bilinear interpolation method,and fuse it with shallow large-scale feature maps to improve the characteristics of shallow layers.Semantic information makes the F-SSRD algorithm improve the detection accuracy of small objects.The F-SSRD algorithm on the NWPU VHR-10 dataset achieves superior results over the current detection algorithm.2.A remote sensing image object detection algorithm based on bilinear interpolation Feature fusion Single-Shot Refinement neural network Detector version 2(F-SSRDv2)is proposed.This algorithm is mainly for the detection MAP of the F-SSRD algorithm.It is hoped that the F-SSRDv2 can further improve the detection MAP of all objects compared to F-SSRD.F-SSRDv2 is an improvement based on F-SSRD.It inherits the detection module of F-SSRD and improves the feature fusion module.It uses a two-step cascaded feature fusion strategy to get better information integration between different hierarchical features.In training,the focal loss function is introduced,and the multi-task loss function is designed based on the focal loss function.The key of this method is to use two cascaded feature fusion strategies to enhance the information combination ability between different hierarchical features,and use the Focal Loss function to avoid the uneven distribution of positive and negative samples in training,thus improving MAP for all objects.The F-SSRDv2 algorithm has a superior effect on the NWPU VHR-10 dataset than the current detection algorithms.3.A remote sensing image terrain classification algorithm based on Deep Learning and Random Forest(DL-RF)is proposed.The algorithm is aimed at the problem that the object categories in the remote sensing image feature classification are different in sparseness degree and the image information is more complicated.Features of different sizes extraction was performed using convolutional neural networks,and fuse these features,and the obtained fusion features were classified using the random forest algorithm.Experimental results on hyperspectral data and LIDAR data show that the DL-RF algorithm achieves superior results over traditional algorithms.