| When applying deep learning to remote sensing ship recognition tasks,a large amount of training set data is usually required to adjust the parameters of the deep neural network.However,since the acquisition of remote sensing ship images usually requires a lot of manpower and material resources,how to target remote sensing ships Recognition of few shot targets in images has become one of the difficult problems.In the few shot target recognition task of remote sensing image,how to solve the problem of insufficient coverage of the weather scene is the key.Secondly,metric learning algorithms are mostly used for few shot target recognition.Compared with general deep learning methods,it solves the problem between few shots and high latitudes,but the recognition speed is also reduced.Therefore,this article conducts research on the above issues.The main research contents are as follows:(1)In view of the problem of incomplete coverage of remote sensing ship image scene caused by insufficient training samples,this paper proposes to expand the training set and support set in different weather scenarios through image style transfer method,and re-establish the relationship network.Train to optimize network parameters.Experiments show that the relational network using style migration has improved the recognition accuracy by about 15% compared with the original relational network.However,the recognition set is augmented due to the use of style migration.In order to solve the problem of slow recognition speed,this paper will use the K-means algorithm to extract the cluster center from each type of remote sensing ship support image high-dimensional feature vector,and extract the extracted high-dimensional feature vector clustering center.The metric learning is performed with the high-dimensional feature vector of the test set image,which greatly reduces the amount of network calculation and ensures the recognition speed.Finally,it is proved by experiments that the relational network with the style migration and the K-means extraction clustering method is added.Compared with the original relational network,the recognition accuracy increases by 13%-18%.(2)For the multi-class remote sensing ship image few shot data set,this paper proposes to use the Part-based Convolutional Baseline(PCB)to classify the multi-category data sets,and then add the style.The overall scheme of migration and K-means extraction clustering center method for fine classification of rough classification samples,and finally proved by experiments that this scheme can make multi-class remote sensing ship image small under the condition that the recognition speed is basically unchanged.The overall accuracy of sample target recognition is increased to about 80%.And based on the above research content,a multi-class remote sensing ship few shot classification prototype platform is realized. |
