Multiscale Deep Feature Fusion Networks For Remote Sensing Image Object Detection And Classification

With the rapid development of modern remote sensing technology,a large number of remote sensing satellites generate massive remote sensing images.Remote sensing images contain very rich information,which has important value in research and application and has been applied to many fields.Remote sensing image object detection and classification are always hot and challenging problems in academic research.In recent years,the development of deep learning technology has significantly improved the performance of remote sensing image object detection and classification.Considering the existing problems of remote sensing image object detection and classification based on deep learning,improved methods for samples selection and networks design are proposed.The main progress in this thesis is as follows:1.An integrated deep convolution fusion network for remote sensing image object detection is proposed.Aiming at the problem of difficult detection of small objects in remote sensing image object detection,K-means-based anchor design and feature fusion are added to the detection network.The size of the initial object proposals is more suitable for the remote sensing dataset,and the extracted features are more suitable for small objects.For the background false detection problem in remote sensing image object detection,the background classification sub-network is added to the detection network.The accuracy of the classification of object proposals is improved by adding the background categories and integrating the classification results.Moreover,by adding this sub-network,the information of the remote sensing image is more fully utilized.The comparison of the experimental results illustrates the superiority of the proposed integrated deep convolutional fusion network.2.A two-path fusion convolution network for multi-source remote sensing image classification is proposed.For the existing remote sensing image classification methods based on convolutional neural network,fixed size is used to get the patches of the pixels.The neighborhood information does not necessarily have positive effect on the central pixel.For this problem,an adaptive area selection based on superpixels is proposed.The spatial distribution of the image is used to determine the area of pixels input to the network.It generates image patches of various sizes to train the network.In order to make better use of the data of multi-source remote sensing images,a two-path fusion convolution network is proposed.In the network,the features of the two kinds of data are extracted separately.The DPN module is used to fuse the two features.The new features are generated under the premise of retaining the original features.Therefore,the robustness and discriminability of the extracted features are good.The comparison of experimental results on different datasets verifies the effectiveness of the proposed method.3.An adaptive multiscale deep fusion residual network for remote sensing image classification is proposed.In order to improve the quality of training datasets,an important samples selection strategy is proposed.Based on the superpixel segmentation results of remote sensing images,gradient information and spatial distribution are used to select diverse and representative training samples.In order to make better use of the hierarchical features in networks,an adaptive multi-scale deep fusion residual network is proposed.Considering that different sizes of targets require different semantic information,an adaptive fusion module is proposed in the network to fuse multi-scale and multi-level features.Comparison of different methods on the four datasets validates the effectiveness of the two innovations.In summary,this thesis studies the remote sensing image object detection and classification.The experimental results demonstrate the effectiveness of the proposed methods and show that the research results have certain practical significance.