High Resolution Radar Detection Based on Neutral Network

Abstract

Neural Network has been broadly applied in the area of modern information technique including radar area. Traditional detection techniques can not detect objects accurately in High Resolution Radar system. Neural Network method can apperceive the tiny diversification of the object because of its flexible parallel processing structure and better mistake toleration. Back-Propagation method solved the problem of multiple network training. But this method is easy to shake and the convergence speed is slow. The Weight adjustment arithmetic based on momentum method can improve in the two aspects. The test shows the efficiency of this kind of adaptive BP method. KEYWORD: High Resolution Radar; Neural Network; Adaptive BP method International Conference on Industrial Technology and Management Science (ITMS 2015) © 2015. The authors Published by Atlantis Press 412 2.2 BP neutral network model The BP arithmetic is divided network learning into two phases. The first phase is inputting known learning sample. The output can be computed by network structure, weight and threshold set before from the first layer to the last layer. The second layer is modification of weight and thresh. The impact (grads) to the whole error can be computed from the last layer to the first layer. The weights and thresh can be modified by grads. The two phases process alternatively to convergence[3]. This method can be spread to network with multiple layers. BP network includes input layer, hidden-layer, and output layer. The model is composed of Input-output model, transfer function model, error computation model and self-learning model[7]. 3 INTRODUCTION OF DETECTION METHOD USING NN The reference [1] proposed the detection principle of using NN. The main idea of the author will be introduced in section 3.1 to 3.3. Section 3.4 gives the method in the case of multiple objects. 3.1 Analyzation of HRR 1D image For LRR (Low Resolution Radar), the receiving signal ( ) R S f (DFT result) is identical with the transmitting signal ( ) T S f .So the matched filter is the optimal filter for LRR. For HRR, the impulse response is ( ) M f . The receiving signal represents in type (2). ( ) ( ) ( ) R T S f S f M f  (2) Then the matched filter characteristic of frequency spectrum is in type (3). 0 2 * * ( ) ( ) ( ) T j ft H f kS t M f e    (3) The matched filter is not only dependent on transmitting pulse but also on the characteristic of the target structure. It is not the best filter in this case. The detection of the 1D(One-dimension) range image is a alternative analytical method[1]. Figure 2 is the 1D range image of one object[4]. Figure 2. 1D range image of HRR echo 3.2 The optimal perceptron neutral detector The aim of radar detection is to make sure the absence of a signal. Hence, the hypothesis testing problem can be stated as type (4). ( ) ( ) ( ) ( ) s t c t x t c t      1 0 H H (4) In type (1), ( ) x t is the whole echo signal. ( ) s t is the object echo signal. ( ) n t is clutter. 1 H means the existence of object. 0 H means the absence of object. The detection problem comes down to the classification of two modes. The two kinds of modes can be implemented by training multiple layers sense sorter[1]. It is assumed that 1D range image of the object and clutter are  , 1,2, , i m s i N  and  , 1,2, , i c c i N  . m N and c N are the number of the object signal sample and the clutter signal sample. ( ) ( ) ( ) 1 2 , , , T i i i i n s s s s      , ( ) ( ) ( ) 1 2 , , , T i i i i n c c c c      (5) In the output end of the network, the anticipant output of the network is 1 m d  and 0 c d  . The target function is defined as type (6).