Multi-Modal ISAR Object Recognition using Adaptive Deep Relation Learning

An efficient complex multi-modal ISAR (Inverse-Synthetic-Aperture-Radar) object recognition method, based on an Adaptive Deep Relation Learning (ADRL), is proposed. (1) Adaptive Dilated Deformable Convolution Neural Network (ADDCNN) is proposed to greatly improve CNNs' multimodal sampling and transformation ability, and increase output feature maps' resolution. Particularly, an adaptive related multimodal regions ranking and selection strategy are proposed to rank and select related regions. (2) DARGL (Deep Attribute-Relation Graph Learning) is proposed to jointly estimate large numbers of heterogeneous multi-attributes, and leverage the inter-modal associations among features and attributes. Extensive experiments are performed on two real-world multi-modal ISAR datasets, showing the proposed method outperforms state-of-the-art methods.