Research On RCS Data Target Recognition Method Based On Meta Learning

In recent years,with the rapid development of radar monitoring and deep learning technology,RCS data target recognition based on deep learning has become a new research direction,which is widely used in military,civil and other fields.In the process of using deep learning technology to train RCS Target recognition model,we often encounter the problem of low target recognition accuracy due to the difficulty of sampling samples and the inability to provide sufficient samples for model training.In order to solve this problem,this paper adopts two different meta learning algorithms from different angles,so that the target object recognition can be realized in the case of a small number of samples.The main research work of this paper is as follows:（1）The common target recognition algorithms with small samples are studied and analyzed.Among them,the network model and training algorithm of MAML algorithm are studied and improved.Firstly,according to the characteristics of MAML’s own shallow network,there are some problems such as weak feature extraction ability and less correlation between data.This paper makes a series of improvements to its network model: using inception module to increase the width and depth of the network and improve the feature extraction ability;Increase the depth of deep output layer and enhance the correlation between neurons;The regularized cross loss function is added on the basis of the original cross loss function L2.Secondly,aiming at the problems of large amount of calculation,gradient disappearance,gradient explosion and unstable training in the process of model training,the training algorithm is improved as follows: the first-order approximation method is used to train the model,and the multi-step loss optimization method is used to update the model parameters.The improved algorithm model has better recognition performance than other models in the case of small samples,and its average prediction accuracy is 84.25%,which is 5 percentage points higher than that before the improvement.（2）Aiming at the problems of weak representation of category center and one-sided selection of measurement function in prototype network,a small sample target recognition algorithm based on measurement adaptation is proposed.Based on the prototype network,the algorithm considers the correlation between query samples and support samples by adding attention mechanism,and then generates a more representative category prototype;Then,feature migration is carried out on the category prototype,so that each category prototype contains inter class information;Then,the task adaptive task embedding network is added to combine the sample category with the task information;Finally,the learnable metric scaling coefficient is added to make the output metric always in an appropriate range.The accuracy of（3）the proposed algorithm for RCS Target recognition is 84.667%,which verifies the effectiveness of the algorithm.To sum up,this paper identifies RCS targets through two different meta learning algorithms,which reflect different ways of thinking and solve the problem of few target samples and difficult training.