Weiliang Huang, and Daqi Zhu
Segmentation network; Underwater environment; Multi-interactive refinement; Salient object detection.
Underwater environment serves as a critical resource for scientific research and marine applications. To suppress noise and gener- ate accurate saliency maps in underwater images, we propose an end-to-end segmentation network. A coarse-to-fine strategy will be adopted which is composed of an extended deep downsampling net- work and a multi-interactive refinement network. The extended deep downsampling encoder will be employed to obtain more pro- found features, which may enhance the representation ability of the network. Noise will be suppressed by the self-refinement module and the attention mechanism in each layer. The extended deep features will be fused to form a coarse saliency map. However, this map may be over-segmented during the noise suppression pro- cess, resulting in a loss of spatial coherence information. The pro- posed multi-interactive refinement network is employed to refine the coarse saliency map, aiming to generate an accurate and complete object representation while minimizing noise introduction. Experi- ments on the underwater datasets and the terrestrial dataset demon- strate that our proposed network achieves accuracy scores of 0.9079, 0.8458, 0.7941 and 0.8757 with the lowest error scores of 0.049, 0.0799, 0.0616 and 0.0353 from the USOD, UFO-120, USOD10K, and DUTS datasets, respectively. The code of our method is avail- able at https://github.com/leckie711/EDMR.
Important Links:
Go Back
