Auto-USOD: Searching Topology for Underwater Salient Object Detection

Tingwei Liu, Runyu Wang, Miao Zhang, Yongri Piao, Huchuan Lu

Abstract: Salient object detection (SOD) methods for terrestrial scene images (TSIs) have obtained remarkable achievements. However, the exploration of SOD for underwater scene images (USIs) is still limited and challenging. Different from TSIs, USIs are usually captured from unconventional shooting angles (overhead, elevated view). Moreover, USIs typically suffer more severe impacts from attenuation, refraction and backscattering, which induces noise interference and further increases the difficulty of mining complete saliency cues. The manually designed CNN-based methods require the extensive experience and intensive labor of human experts, not flexible enough to cope with these factors for underwater SOD. To this end, we propose an end-to-end network named Auto-USOD for underwater SOD, including a topology search scheme and a search space. The search scheme allows the network adaptively coordinates the cooperation among feature extraction, feature selection and feature fusion. Moreover, the search space is divided into encoder and decoder search spaces, which enable the network to sense and recover features of saliency objects in the encoder and effectively fuse the fine-coarse features in the decoder, respectively. Extensive experiments on two benchmark underwater datasets demonstrate the effectiveness of the proposed method and achieve comparable performance on four evaluation metrics. Note that our model is 2.3 Mb, only 3.1% of second-best model in parameter size and less than 87% in Flops, while our model outperforms it in terms of performance. The source code will be publicly available at https://github.com/LiuTingWed/Auto-USOD.