Research On Remote Sensing Scene Classification Algorithm Based On Double Branch Network

The goal of remote sensing scene image classification is to assign specific semantic labels to remote sensing images,which is significance and foundation in many practical applications.Recently,with the development of deep learning models,convolutional neural networks have shown powerful feature extraction capabilities in combining local and global features,and deep learning methods have been widely used in remote sensing scene image classification.Although these deep learning models have greatly improved the performance of remote sensing scene image classification,the problems of confusing images that are difficult to discriminate,differences in image scales,and complex backgrounds in remote sensing images are still a great challenge for remote sensing scene image classification.To address the above problems,this paper proposes two remote sensing scene classification algorithms based on double branch network to learn stronger feature representation and extract more discriminative features.The main work is as follows:1)A remote sensing scene image classification algorithm based on comparison double branch network is proposed.To solve the problem of indistinguishable confounded images,the proposed method aims to distinguish confounded images by learning the focus area of the input image pair and the difference area between two images.The main steps of the algorithm are as follows: first,image features are extracted using pre-trained Res Net50,and the most similar pairs of images are selected by the Euclidean distance between the features as the input of the two branches.Then the image pairs are expressed in pairs: by introducing self-representations to highlight the informative part of each image itself,and by introducing mutual-representations to capture the subtle differences between image pairs.Third,the obtained representations are used to predict the class of the input images,while adding ranking loss enables the network to adaptively enhance the recognition of features with different expression priorities.The algorithm achieves 95.53% as well as 96.76% overall accuracy at 20%and 50% of the training data on the AID dataset.It also achieves 92.64% as well as94.59% overall accuracy at 10% and 20% of training data on the large remote sensing scene classification dataset NWPU-RESISC45 dataset,which is an improvement over the existing methods.The comprehensive experimental results demonstrate that the proposed method can effectively discriminate similar images by focusing on the subtle differences of images,thus improving the capability of the network in remote sensing scene classification as well as its application value.2)A remote sensing scene image classification algorithm based on local-global double branch mutual learning network is proposed.The proposed method mainly captures global and local features for problems such as inconsistent image scales and complex backgrounds,which lead to a small percentage of discriminative regions that are not obvious.The algorithm has two branches,the global branch and the local branch,respectively.Specifically,the global branch uses a Convolutional Neural Network to construct a heat map from the input image: high activation semantic regions and low activation irrelevant background regions.The resulting heat map is then multiplied with the original input image to obtain local regions.These local regions contain the most scene recognition regions,and the features extracted from these regions are more focused on differentiating the scene representation.Next,mutual learning is introduced to learn the correlation between local branches and global branches using the feature differences between the original image and local regions.The local-global branches are also trained using classification loss and mutual learning loss to constrain the two branches to promote each other.The algorithm is validated on two datasets,the overall accuracy of 96.21%(20% for training data)and 97.86%(50% for training data)is achieved on AID data set,the overall accuracy on NWPU-RESISC45 dataset is 92.67%(training data is 10%)and 94.73%(training data is 20%).both achieving the best accuracy on both datasets.The experiments show that the algorithm can effectively improve the correct classification rate of remote sensing scene images by using the prediction differences between global and local branches to facilitate network training through the mutual learning mechanism.Meanwhile,the algorithm achieves the best performance on multiple datasets,indicating its good model generalization ability.