Open Set Recognition Of Fine-Grained Remote Sensing Ship Images Based On Decision Fusion

Ships play an important role in modern seas.Militarily,accurate ship identification can enhance coastal defense and ocean-going strike capabilities.On the civilian side,ship identification technology can help relevant authorities conduct maritime search and rescue and traffic control activities.Therefore,accurate and efficient identification of ship targets is a very important application value in intelligent Marine monitoring system.However,the recognition process of traditional methods is complex and the accuracy is not ideal.At present,the fine-grained classification of remote sensing ships based on deep convolutional neural network is completed under the closed set assumption,which is not conducive to practical application.To solve the above problems,this thesis introduces the open set recognition method into the fine-grained ship image classification,and designs an open set recognition method for fine-grained remote sensing ship images based on branch fusion.The work done is as follows:(1)In order to highlight the target subject of the ship image,the Spatial Transformer Network(STN)is added before the feature extraction network to filter the interference of background information,and a multi-scale parallel convolution structure is splicing after the STN module to extract rich features.Enhance the feature extraction ability of the network for local regions with different scales.(2)For the large intra-class difference of fine-grained remote sensing ship images,this paper uses the center loss function in the base branch to reduce the intra-class difference.In order to get rid of the limitation of traditional image classification based on the closed set assumption and the dependence on the training data set,and make it more suitable for the actual application scenario,the fine-grained problem of ship image is pushed to the open world.Based on the class structure analysis method and visual concept,this thesis designed a meta-embedding branch to strengthen the model’s learning of small tail samples and increase the difference between classes of ship images.In this branch,cosine classifier was used to replace the traditional classifier,so as to identify the known and unknown classes in a unified manner.Finally,the classification results of the two branches are fused and the threshold is set.To verify the effectiveness of the proposed algorithm,a series of comparative tests were conducted on the processed FGSCR-42 published fine-grained remote sensing data set.The experimental results show that the algorithm model in this thesis has a high accuracy under corresponding different openness and is almost unaffected by the data distribution,which proves the feasibility and advantages of the open set recognition method and fine-grained image recognition network.