A Joint Unsupervised Super-resolution and SAR Object Detection Network
Chong-Qi Zhang, Zi‐Wen Zhang, Ming-Zhe Chong, Yunhua Tan · 2023
Synthetic aperture radar (SAR) is an active microwave coherent imaging technology, which can provide images showing scattering properties of targets for target recognition regardless of adverse light and weather conditions. However, the low-resolution and speckle noise problems restrict the accuracy of object detection. Hence, super-resolution (SR) algorithms, providing additional details of targets and suppressing noise, are designed to guide the detection network. Nevertheless, how to effectively combine the two tasks turn out to be a challenge, i.e., the double parameters produced by the extra task consume more computing capacity and memory and thus restrict the interaction speed. Besides, the improper guidance of super-resolution also limits the detection accuracy. In this paper, a joint unsupervised super-resolution and SAR object detection network is proposed. First, a mixed loss derived from two SAR SR metrics replaces the original enormous SR dataset, thereby reducing the computational requirement for an unsupervised SR network. Second, an epoch-aware training strategy is applied in the whole training process to balance the two tasks and prevent overfitting for one of the two tasks. Experiment results on SSDD and MSAR demonstrate the enhancement of the proposed method in terms of average precision with less computing memory.