| With the advancement of remote sensing technology,the emergence of a large number of high-resolution remote sensing image resources has made it possible for people to carry out various detection activities on the earth’s surface.In recent years,with rapid growth of resolution of remote sensing images,traditional remote sensing images at the pixel level and object level cannot meet people’s needs.The reason is that: with the improvement of the spatial resolution of remote sensing images,each pixel content and the distribution of ground objects become more and more complex,making the scene-level classification of remote sensing images has become an urgent problem to be solved.Based on this background,this thesis conducts research on the performance improvement of the scene classification method for high-resolution remote sensing images based on deep learning from the following two aspects.1.The method of remote sensing scene classification based on metric learning.In this thesis,a scene classification network based on metric learning is constructed,and the multi-level features of anchor samples,positive instance samples and negative instance sample images are respectively extracted through the siamese network architecture.It’s goal is to reduce the distance between all anchor samples and negative instance sample features,with increasing the distance between all anchor samples and negative instance samples.In order to balance the details and semantic information in the feature map,feature pyramid network is introduced to effectively solves the problem of semantic ambiguity caused by the high intra-class difference and big inter-class similarity of remote sensing images.In addition,in order to strengthen the supervision to the network,a joint loss function is introduced to take the utilization of training supervision information into account.It is experimented that the algorithm can obtain better scene classification results.2.The method of remote sensing scene classification based on enhanced multi-scale and global information.Since the convolution operation ignores the global information of the remote sensing image,it is difficult for the scene classification method using CNN architecture to distinguish some categories with complex scene distribution.In order to solve the mentioned problems,a Transformer network codec module is introduced,inspired by the idea of patch-embedding.Then the thesis establishes global context information associations for regions of different sizes in remote sensing images.In order to integrate multi-scale feature information,a Transformer multi-scale global information enhancement module is introduced to fuse multi-scale feature encodings,realizing to get multi-scale semantic modeling of feature embedding and reduce global information waste.It is experimented that the algorithm is able to learn more robust images features to improve its scene classification effect.3.The method of remote sensing scene classification based on hybrid network architecture.Although the deep convolutional neural network architecture has certain advantages,its receptive field is usually small,which is not conducive to capturing global features.And because the visual Transformer architecture can capture the global information of a picture,it has achieved better results than the convolutional neural network architecture in many visual tasks.Inspired by the idea of integrated networks,the two network architectures of convolutional neural network and Transformer network are integrated,and the self-attention mechanism is integrated into the convolutional neural network architecture to realize the global improvement of low-level network features with a relatively small amount of calculation.Completion,and the global feature association ability of advanced features,thereby improving the network’s ability to recognize scale-changing objects and complex distribution scenes. |
