Continual Learning for SAR Target Incremental Detection via Predicted Location Probability Representation and Proposal Selection
Yu Tian,Zongyong Cui,Jizhen Ma,Zheng Zhou,Zongjie Cao
DOI: https://doi.org/10.1109/tgrs.2024.3394157
IF: 8.2
2024-05-14
IEEE Transactions on Geoscience and Remote Sensing
Abstract:The gradual increase of synthetic aperture radar (SAR) imagery often accompanies the appearance of new targets, but traditional detection frameworks can only detect existing target classes and cannot detect new ones. Typically, we must update the model using both new and old data, which puts a strain on storage and computation, but if we only update with new data, the detection performance on old classes will suffer dramatically. For this reason, this article proposes to use the continual learning (CL) method to solve the problem of SAR target incremental detection. Mainstream CL methods generally consider localization to be a class-irrelevant function, however, this strategy is unsuitable for SAR imagery with significant background changes, leading to poor detection performance. Addressing the above issues, this article proposes a CL object detection (CLOD) method, with an overall framework based on knowledge distillation. The focus of the methodology consists of two parts: first, we introduce the predicted location probability representation (PLPR) method, by using spatial discretization and segmented probabilistic statistics, to transform the localization results into probability distributions, thus allowing the localization function to participate in the CL process; second, we design a proposal selection strategy according to the background characteristics of SAR images, which improves the quality of the proposals during the knowledge review to further optimize the learning effect. Experiments on the latest multiclass SAR target detection dataset MSAR-1.0, show that our method is able to learn new knowledge with less performance penalty for old classes than other methods. In multiple data incremental settings, our method provides a 2%–11% performance improvement over numerous common methods.
engineering, electrical & electronic,imaging science & photographic technology,remote sensing,geochemistry & geophysics