SSoB: searching a scene-oriented architecture for underwater object detection

Wanqi Yuan, Chenping Fu, Risheng Liu, Xin Fan

Abstract: Underwater object detection (UOD) suffers from low detection accuracy because of environmental degradations, such as haze-like effects, color distortions, and imaging noises. Therefore, we commit to resolving the issue of object detection with compounded environmental degradations that greatly challenges existing deep learning-based detectors. We propose a neural architecture search -based deep learning network to realize the UOD task, which can automatically discover the scene-oriented feature representation. Our network is accomplished through a unified macro-detector and a novel mixed anti-aliasing block (MAaB)-based search space. The macro-detector targets to learn intrinsic feature representations automatically from underwater images containing various environmental degradations and complete the subsequent detection tasks. The novel MAaB-based search space is proposed toward complex underwater scenes. The candidate operator MAaB has multiple kernels and anti-aliased convolutions in a single block for boosting the contextual representation capacity and the robustness of degraded factors. Finally, we use the differential search strategy guides the whole learning process to obtain the scene-friendly results. Extensive experiments demonstrate that our method outperforms the state-of-the-art detectors by a large margin. More importantly, in cases where environmental degradation is severely disturbed, our method is also superior to other popular detectors.