Object Recognition In Remote Sensing Image Based On Feature Balance And Diversity

Remote sensing image object recognition is an image-level task,the purpose of which is to distinguish the categories of objects in remote sensing images,and it is widely used in military security,maritime traffic and so on.At present,using attention mechanism to obtain identifiable features and adding detection,segmentation networks and other sub-networks to find the key areas needed for recognition are the main research methods of remote sensing image object recognition.However,most of the research on object recognition in remote sensing images is only applicable to high-resolution remote sensing images.This thesis proposes an object recognition method based on feature balance,to solve the problem of low recognition accuracy in low-resolution remote sensing images.Since the low-resolution images contain limited details,this thesis introduces the super-resolution task as an auxiliary task to provide details that is beneficial to recognition.In addition,ordinary feature fusion methods cannot effectively utilize the features provided by super-resolution tasks.In order to solve this problem,a remote sensing image recognition method based on feature balance is proposed,which selectively incorporate strong and weak features in the features of the super-resolution task into the coarse feature of low-resolution image.Then,optimizing the super-resolution branch using gradient-weighted class activation maps generated by the classification branch.The above process goes through multiple interactive iterations to improve the classification accuracy.To facilitate research on low-resolution remote sensing images,this thesis constructs and discloses a new dataset named LFS,which contains 4591 low-resolution remote sensing images along with labels.In addition,remote sensing images are difficult to obtain and the cost of labeling is high,and the labeled data for recognition is limited.In order to solve the problem of a small amount of labeled data,this thesis also studies how to use limited training data to obtain a recognition network with high recognition accuracy,and proposes a remote sensing image object recognition method based on negative correlation loss.Since ensemble learning can increase the diversity information in the network by generating multiple individual learners,and improve the overall performance of the network,this thesis use the ensemble network as the basic architecture.In order to effectively take advantage of the ensemble network,this thesis designs a negative correlation loss,adds feature constraints between different sub-models,and reduces instances of class boundary errors caused by overfitting to a small number of training data.On this basis,using the average strategy to integrate multiple outputs of the network.In this way,the error of diversity results cancel each other out during integration,thereby improving the recognition accuracy of the model.In summary,from the perspective of low-resolution remote sensing images and a small amount of label data,this thesis proposes remote sensing image target recognition based on feature balance and remote sensing image target recognition based on negative correlation loss.The experimental results show that the recognition method based on feature balance can effectively recognize the objects of low-resolution remote sensing images,and the remote sensing image object recognition network based on negative correlation loss can achieve high recognition accuracy under the training of a small number of samples.