DIOD: Fast and Efficient Weakly Semi-Supervised Deep Complex ISAR Object Detection

Abstract

Inverse synthetic aperture radar (ISAR) object detection is one of the most important and challenging problems in computer vision tasks. To provide a convenient and high-quality ISAR object detection method, a fast and efficient weakly semi-supervised method, called deep ISAR object detection (DIOD), is proposed, based on advanced region proposal networks (ARPNs) and weakly semi-supervised deep joint sparse learning: 1) to generate high-level region proposals and localize potential ISAR objects robustly and accurately in minimal time, ARPN is proposed based on a multiscale fully convolutional region proposal network and a region proposal classification and ranking strategy. ARPN shares common convolutional layers with the Inception-ResNet-based system and offers almost cost-free proposal computation with excellent performance; 2) to solve the difficult problem of the lack of sufficient annotated training data, especially in the ISAR field, a convenient and efficient weakly semi-supervised training method is proposed with the weakly annotated and unannotated ISAR images. Particularly, a pairwise-ranking loss handles the weakly annotated images, while a triplet-ranking loss is employed to harness the unannotated images; and 3) to further improve the accuracy and speed of the whole system, a novel sharable-individual mechanism and a relational-regularized joint sparse learning strategy are introduced to achieve more discriminative and comprehensive representations while learning the shared- and individual-features and their correlations. Extensive experiments are performed on two real-world ISAR datasets, showing that DIOD outperforms existing state-of-the-art methods and achieves higher accuracy with shorter execution time.