Research On Fine-grained Classification Of Remote Sensing Ship Images Based On Attention Mechanism

Fine-grained classification is an important and challenging research direction in the field of computer vision.Compared with coarse-grained classification,its difficulty lies in the small inter-class difference and large intra-class difference.Fine-grained classification of natural images has made significant progress with the release of popular datasets such as CUB-200-2011,Stanford Cars,and Aircraft datasets.However,in the field of remote sensing images,affected by the lack of publicly available datasets,there is still a large room for improvement in the current fine-grained ship classification research.The accurate classification of ships on the sea surface is an important part of maritime monitoring,which is of great significance in civil,commercial,military and other aspects.In recent years,the attention mechanism has shown excellent expressiveness in the field of computer vision,which can help the neural network to better locate the discriminative area in the image,and improve the discrimination ability and robustness of the network,which has a good application value in the actual classification scene.Based on this,this thesis proposes a finegrained classification method of remote sensing ship images based on attention mechanism,and proposes corresponding solutions to some main problems by using core technologies such as attention mechanism and multi-granularity feature interaction.The main research contents are as follows:1.The challenge of fine-grained classification of remote sensing ship images lies not only in the difficulty of discriminative region localization,but also in the scarcity of publicly available datasets.Therefore,this thesis proposes a fine-grained classification method for remote sensing ship images based on mixed attention mechanism.Firstly,the mixed attention mechanism CBAM was used to generate the attention map of each training image to highlight the salient feature parts of the target.Secondly,data augmentation was performed by attention-guided region clipping and attention-guided region deletion.Finally,the original image and the enhanced image are used as input for training.The proposed method is verified on the datasets FGSCR-42 and FGSC-23.Experimental results show that the proposed method can effectively improve the accuracy of fine-grained classification of remote sensing ship images.2.Most of the current fine-grained classification works based on convolutional neural networks use feature maps extracted from the last convolution layer to mine discriminative regions.However,the large receptive field makes the last convolutional layer tend to focus on the whole object and ignore many detailed features,which leads to a decrease in the ability of the network to identify differences.In order to solve the above problems,this thesis proposes a fine-grained classification method for remote sensing ship images based on multigranularity feature interaction.Firstly,an effective cross-layer trilinear pooling was introduced to calculate the third-order interaction between three different layers.The thirdorder interactions of different combinations are then fused to form the final feature representation.In addition,to further improve the performance,an attention module is introduced,which enables the model to effectively locate the local area of the target given the original image.The proposed method was verified on the remote sensing ship fine-grained classification datasets FGSCR-42 and FGSC-23.Experimental results show that the proposed method has better classification performance than most mainstream fine-grained classification methods.