Semantic enhanced image blending for spaceborne SAR airport targets
Yi Kuang, Yingbing Liu, Wei Hu, Xiaoming Xie, Fan Zhang · IET conference proceedings. · 2024
As the extensive application of deep networks and the increasing complexity of functionalities, there is a growing demand for training samples. In most publicly available SAR target detection datasets, the number of target instances is limited. Theref ore, SAR sample synthesis becomes an effective approach to enhance the target detection capability of deep models. In the task of image synthesis, image blending aims to smoothly integrate objects from source images into other background images in a natural manner. This paper proposes a semantic-enhanced image fusion method for achieving high-fidelity and reliable sample augmentation. Firstly, an object feature separation method is introduced, which separates the target features into key textur es and secondary content. Secondly, Poisson blending and style transfer are applied separately to these two feature parts, with controllable degrees of feature fusion achievable through different parameter settings. Finally, deep fusion is performed on the entire image to further enhance the appearance realism of samples. Experimental results demonstrate that semantically enhanced fusion images can be used for sample augmentation and significantly improve the recall rate (R) of classical detection networks.