Composite Data Preparation Algorithm for SAR Imagery Object Recognition

The paper presents the approach to resolve the insufficient training data problem for object recognition on radar images. Article proposes the novel strategy of data preparation based on the fusion of objects and background images to form composite data. The composite samples were formed by the Poisson's image blending algorithm using open-source experimental database MSTAR and data synthesized by the pix2pix algorithm. It was shown that composite data is applicable in the conditions of a low number of initial data for training object detectors such as YOLOv5. According to mAP for the considered insufficient amount of training data case reaches 82%. It is shown that implementing training strategy based on the prepared composite dataset improves the accuracy recognition up to 3% of mAP. Generalization ability of trained detector was shown on delayed manually annotated samples from ICEYE database where model achieved mAP@0.5 of 62%.