Location Distributed Modeling for Improving Scale Adaptability of SAR Incremental Target Detectors

With the continuous increase in Synthetic Aperture Radar (SAR) data, it becomes a real challenge to efficiently update existing SAR target detection models. The incremental target detection method is a solution that can update the model using only new data while preventing old knowledge from being forgotten. However, existing incremental target detection methods lack consideration of the characteristics of SAR targets, which does not lead to desirable results in SAR target detection. There may be significant scale differences between targets of different SAR classes, and if cross-scale targets participate in the incremental learning process, the localization ability of the detector will be greatly reduced. For this reason, Location Distributed Modeling (LDM) is proposed in this paper to improve the scale adaptability of SAR incremental target detectors. The method first transforms the traditional bounding box into a rectangular distribution by 2-D Gaussian modeling and further forms a new incremental localization loss by measuring the difference between the distributions. LDM can weaken the sensitivity of the incremental target detector to the scale of the target and improve attention to the center offset and the overall distribution. We set up several data incremental scenarios with the multi-class SAR target detection dataset MSAR in this paper, and the experimental results show that our method significantly improves the scale adaptability of SAR incremental target detectors.