Towards salient object detection via parallel dual-decoder network

Abstract

Salient object detection, an important preprocessing step in computer vision, segments the most prominent objects in an image. However, existing research in this field utilizes transformer-based methods to capture global context information, failing to effectively obtain local spatial features. To solve this issue, we propose a parallel dual-decoder network, which consists of a novel semantic decoder and a modified salient decoder. Specifically, the proposed semantic decoder is designed to learn the local spatial details, and the salient decoder utilizes the learnable queries to establish global saliency dependencies among objects. Moreover, the two decoders establish correlations between saliency and multi-scale semantic representations through cross-attention interaction, significantly enhancing the performance of salient object detection. In other words, we obtain global context information in the decoder to prevent discriminative features from being diluted during information propagation. Extensive experiments on 15 benchmark datasets demonstrate that our model significantly outperforms other comparison methods and shows promising potential for real-world applications such as challenging optical remote sensing, underwater, low-light, and other open scenarios. In addition, our method shows excellent performance in other downstream tasks such as camouflaged object detection, transparent object detection, shadow detection, and semantic segmentation.