Hierarchical Space-Channel Context Correlation for Salient Object Detection

Salient object detection (SOD) plays a critical role in image-based measurement systems by identifying the most visually conspicuous objects in an image. Despite significant progress due to the strong feature representation capabilities of neural networks, current methods tend to focus on either spatial context modeling or channel context modeling, often overlooking the comprehensive correlations between them. In this work, we thoroughly explore the space-channel context correlation within an image scene to derive more effective saliency representations. Specifically, we propose a hierarchical context correlation network (HCC-Net) that leverages hierarchical contextual information, i.e., scene-, region-, and pixel-level contexts, to establish comprehensive space-channel correlations. These three types of context correlations are unified through a bilinear correlation approach, which is the outer product of two vectors representing space and channel descriptors. This method is both easy to implement and computationally efficient, capturing complex space-channel dependencies within an image and enhancing the accurate localization of salient objects. Additionally, to address the misalignment issue that can arise among different network levels, we design a high-level semantics guided feature alignment module (HSA). This module aligns representations at each level with high-level semantic representations through an efficient cross-attention mechanism. Extensive experiments on five widely used benchmarks demonstrate that our proposed method performs favorably against the state-of-the-art methods. Importantly, the evaluations highlight the significance and efficacy of comprehensive space-channel correlation for SOD.