Research On Target Detection Method Of Rader RD Images Based On Deep Learning

Due to the time-consuming processing required to generate synthetic aperture radar（SAR）images,target detection based on SAR images is usually not suitable for real-time applications.Therefore,direct use of radar Range Doppler（RD）data for target detection can quickly detect the enemy target location information,which plays a very important role in the military battlefield.There are two difficulties in target detection based on radar RD data: first,for such task,the traditional CFAR algorithm are not universal.Once the distribution of background noise changes,it is necessary to reselect parameters to build a mathematical model.Under the requirements of high-precision detection,a slight deviation in the detection will be judged as false detection.Second,in the combat environment with high mobility and strong interference,target occupies a small proportion of units in the generated radar RD images,and is affected by false targets and strong noise,which will lead to the inability to correctly identify the real target and the detection performance will reduce.At present,some frequently-used methods based on deep learning can’t meet the requirements of accurate positioning when dealing with such problems,and there will be missed detection and false detection in the case of complex scenes.Based on radar target detection and neural network theory,this thesis focuses on target detection of radar RD data and accurate target detection in complex scenes.In this thesis,the basic structure of CNN and the formation principle of RD spectra are introduced in detail,and data simulation experiments are carried out according to the basic knowledge of radar signals.Aiming at the problem of target detection based on radar RD data,this thesis adopts a cascade neural network structure.The backbone network uses Res Net50 to extract features.By cascading layer by layer,the over-fitting problem caused by directly increasing the training intersection over union threshold is effectively alleviated,and ensures the number of positive samples in training,so as to obtain more and more highquality proposal boxes and effectively improve the detection accuracy.Aiming at the problem of radar target detection in complex environment,this thesis first improves the regression loss function,obtain high-quality proposal boxes by measuring the distance between the center of the proposal boxes and the real boxes.The smaller the distance,the better the quality of the proposal boxes.Then,because the target occupies a small proportion of units in the RD images,a hybrid attention module is added in the shallow part of the backbone network to extract better semantic features,so as to reduce the false positives caused by some noise and interfering targets.For the target detection problem of radar RD images,the performance of the detection method of cascade neural network proposed in this thesis is better than the traditional CFAR and yolo v3 algorithm.For targets in complex scenes,this thesis further improves the cascaded neural network,and the detection effect is also improved.Experiments show that this method is more accurate than other existing methods.