Research On Matching Area Selection Of SAR Image Based On Incremental Learning

As The selection of SAR scene matching area is a key technology of unmanned aerial vehicle navigation planning.Using guarantee data to select areas with excellent matching performance is of great significance to improve the success rate of scene matching.Compared with the traditional rule-based matching area selection method,the matching area selection technology based on deep learning has stronger feature self-learning extraction and matching modeling capabilities.After the model is trained,the matching area can be automatically selected and become the matching area selection The main direction of development.The performance of the deep learning model is closely related to the training samples,and misselected problem may occur in new task scenarios that are not covered by the training data.Therefore,when the training samples of the original model cannot be obtained,how to update or even correct the misjudgment of the selected model,and avoid the catastrophic forgetting problem caused by the update of network model parameters,is a new approach in the study of the knowledge increment method for the selection of the adaptation area.challenge.The SAR scene matching area is an unstructured object,and single-scale features are difficult to describe its adaptability.Aiming at the problem of changing sample patterns in the SAR scene matching area,a two-stage adaptation area detection method based on Feature Pyramid Network(FPN)is designed.Multi-scale features are integrated to realize multi-level classification and positioning of objects in the matching area.Aiming at the problem of missing training samples for old tasks faced by incremental matching area selection,an incremental selection method of matching area based on feature extraction incremental feature pyramid network(IFE-FPN)was designed,and the old shared layer parameters were retained by the fixed network model.Knowledge,independent training of the proprietary layer of the new category,to maximize the prediction performance of the old task,while achieving the knowledge increase,and the error correction of the wrongly selected samples in the way of category increment.Aiming at the problem that the feature extraction method has limited ability to distinguish new categories,a matching area model update method based on the incremental feature pyramid network(IKD-FPN)of knowledge distillation is designed,and the old knowledge is retained by introducing the distillation loss,and the data of the new category is updated at the same time.The model parameters effectively improve the error correction capability of the non-adapted area category,and also overcome catastrophic forgetting.Finally,the performance of the proposed incremental selection method and the non-incremental model,fine-tuning method,and joint training method are compared and analyzed.The experimental results prove that the incremental selection methods can achieve effective knowledge increment,thereby eliminating misselected samples.Improve the performance of scene matching area selection.