Research On Sar Image Target Recognition Technology Based On Deep Learning And Decision Tree

Synthetic Aperture Radar(Synthetic Aperture Radar,SAR),because of its unique characteristics,using the echo signal to collect images so that it can detect certain penetration without interference from external conditions.So it has very practical application and great using value.Due to its high cost of image interpretation,SAR automatic target recognition research has great value.With the deepening of research,the current research on traditional methods has caused SAR target recognition to encounter a bottleneck.The application of deep learning in various fields in recent years has also made better results in the field of SAR target recognition.However,due to the limitation of data volume,under the condition of small samples,SAR target recognition based on deep learning has encountered new challenges.In order to enhance the method of target recognition,we solve the problem of limited samples from two aspects of feature enhancement and integrated learning strategy,and further study the target recognition method of SAR images based on deep learning and decision tree.The main contents of this article are as follows:First,after in-depth analysis of the traditional SAR image target recognition method and its shortcomings,the deep learning method is applied to SAR image target recognition.Then it analyzes the deficiencies of deep learning schemes under the condition of limited samples.Then,under the restriction of small samples,in order to enhance the feature extraction method,the method of multi-layer cascade network is introduced into SAR target recognition.Generally,high-level features are more comprehensive and distinguishable than low-and medium-level features,and are often used for category recognition.In order to make up the shortage of training features in the case of limited samples,the concatenated features of the optimally selected convolutional layer are concatenated to provide more comprehensive features for recognition.Therefore,the performance of SAR target recognition under the condition of small samples is better improved.Finally,in order to provide a more powerful target recognition scheme,the idea of integration is introduced for the basis of cascaded network features,and a SAR target recognition method based on decision tree and cascade CNN is further proposed.In order to make full use of these cascaded features,the classifier AdaBoost rotating forest based on integrated learning is introduced to replace the original softmax layer to achieve more accurate limited sample recognition.This method effectively enhances the ability of target recognition from the level of features and classification methods.Through the AdaBoost rotation forest method,not only the rotation matrix is used to further enhance these features,but also the weak classifier with different weights is used to construct a strong classifier.