TAFDet: A Task Awareness Focal Detector for Ship Detection in SAR Images

Yilong Lv, Min Li, Yujie He, Yu Song

Abstract: With the wide application of Synthetic Aperture Radar(SAR) radar in maritime surveillance, ship detection method has been developed rapidly. However, there is still a key problem that is the general parallel detection head network has task inconsistency between different tasks. In this work, we proposes a task awareness focal Detector (TAFDet) that alleviate the problem of task inconsistency, consisting of two sub-networks: task awareness attention(TAA-Subnet) and cross-task focal sub-network (CTF-Subnet). Firstly, we design TAA-Subnet that can use the standard backpropagation operation to transform the gradient signal into an attention map of different tasks. To select prominent attention features and obtain a task awareness mask, we proposed a mask generation module(MGM) in TAA-Subnet. Second, CTF-Subnet is designed to tune features in different tasks using task awareness mask, where a feature focal module (FFM) was proposed to enhance the features not paid attention to in the task. Experiments show that our method consistently improved state-of-the-art detectors on SSDD and HRSID datasets. Specifically, our TAFDet consistently surpasses the strong baseline by 0.4 $$\sim $$ 5.8 AP with different backbones on SSDD and HRSID datasets.