Research On SAR Image Classification Method Based On Multi-azimuth Informatio

Synthetic aperture radar(SAR),whether airborne or spaceborne,can operate day and night in various climates and produce high-resolution images.These images find wide application in key scene surveillance.The SAR Automatic Target Recognition(ATR)algorithm has significant advantages in SAR image interpretation since it can automatically detect the target area of interest and categorize the target from complex images.SAR images are azimuthsensitive,meaning target features change significantly with the change in azimuth.However,the current SAR-ATR algorithms,based on deep learning,do not take the azimuth information of SAR image targets into consideration.Multi-azimuth SAR images provide more comprehensive feature information for target recognition,which is helpful to improve target recognition.Thus,incorporating SAR image azimuth information into the SAR-ATR algorithm holds great research value.This paper aims to address these issues and carries out the following researches,mainly including:(1)Facing with the problems of low accuracy,poor efficiency and insufficient use of azimuth sensitivity in traditional SAR image classification algorithms,a SAR image classification learning method based on azimuth information is proposed.Prior to classifying SAR images,the azimuth Angle of SAR image target is estimated,and input it into the Convolutional Neural Network(CNN)submodel based on SAR target recognition network at a specific angle.Following convolution pooling and other operations,a Softmax classifier is used to determine the target category and obtain the final classification result.By analyzing and comparing several groups of experiments,the classification results demonstrate the effectiveness of azimuth information in SAR-ATR.Furthermore,more precise azimuth information can improve the ability of the convolutional neural network to accurately classify and recognize SAR images.(2)To enhance the accuracy of SAR target classification,a SAR image classification method is proposed based on a multi-azimuth multi-input network,taking into account that multiple azimuth SAR images of the same target contain richer classification and recognition information.In this model,SAR images captured at different azimuth angles are simultaneously input into the convolutional neural network.Two learning methods of sequential progressive fusion and simultaneous fusion are proposed in the multi-input mode.Moreover,training samples that satisfy the multi-input mode could be generated by a multi-azimuth SAR image sample sequence combination method.Through the analysis of multiple simulation experiments with two input and four input in two network modes,it can be seen that the superiority of the proposed algorithm in terms of target recognition performance over other algorithms.(3)The software development of SAR image classification method under the multiazimuth model was achieved using the Python software platform.Firstly,an interactive interface was designed to include input and display controls for test samples,azimuth information,and various image classification algorithms.Next,the image classification algorithms’ related models were tested by making programming calls.Lastly,the software displayed the image classification results,target recognition accuracy,and confusion matrix in the result display area to observe and evaluate the effect of image classification.This process completed the development of the test software for the SAR image classification and recognition method under the multi-azimuth model.