Research On Matching Area Selection Of SAR Image Based On SA-HRNet

The technology of automatic selection of the matching area is one of the key technologies for image matching.Modern military systems usually combine inertial navigation system and synthetic aperture radar scene matching technology because SAR has full-time,all-weather and high-resolution imaging characteristics.Selecting a portion region with high matching accuracy is a prerequisite for accurate guidance.However,the matching area is not clearly defined,which is different from typical targets.The structure of the matching area is diverse,so it is difficult to describe their characteristics with uniform rules.There are still great challenges in matching area selection.This paper focuses on the research of SAR matching area selection.First,a hierarchical rule-based method is introduced to make multi-dimensional judgments in explicit features such as the information complexity and stability of the scene.The inherent speckle noise of SAR image interferes characteristic analysis,so a cascade network is proposed to improve the image quality,which increased the accuracy of matching area selection.In order to enhance the robustness of matching area selection model,this paper proposes SA-HRNet based on convolutional neural network.The SAR images are subdivided according to typical scenes because of the difference in adaptability of different scenarios,matching area selection is converted into a multi-classification problem;and the utilization of lightweight attention mechanism in the network improves the model performance.The method based on deep learning is prone to catastrophic forgetting problems when updating models for new scenes without relying on old data.Therefore,this paper proposes to transfer the old model by means of knowledge distillation,which can maintain the existing classification performance.Meanwhile the weight of easily-separated samples is reduced to solve the problem of imbalanced classes in SAR image datasets,which accelerated the convergence speed of the network.Finally,the proposed model in this paper is compared with other methods on real SAR image data,and the actual availability of the matching areas is verified through simulation experiments.Experimental results show that proposed SA-HRNet is superior to other methods in terms of selection performance,and has good generalization ability and robustness.