Abstract
UNet-based methods have shown outstanding performance in salient object detection (SOD), but are problematic in two aspects. 1) Indiscriminately integrating the encoder feature, which contains spatial information for multiple objects, and the decoder feature, which contains global information of the salient object, is likely to convey unnecessary details of non-salient objects to the decoder, hindering saliency detection. 2) To deal with ambiguous object boundaries and generate accurate saliency maps, the model needs additional branches, such as edge reconstructions, which leads to increasing computational cost. To address the problems, we propose a context fusion decoder network (CFDN) and near edge weighted loss (NEWLoss) function. The CFDN creates an accurate saliency map by integrating global context information and thus suppressing the influence of the unnecessary spatial information. NEWLoss accelerates learning of obscure boundaries without additional modules by generating weight maps on object boundaries. Our method is evaluated on four benchmarks and achieves state-of-the-art performance.
Introduction
Salient Object Detection (SOD) aims to identify and segment the most visually prominent objects in an image, mimicking human visual attention. While U-Net-based architectures are standard for this task, they face two significant challenges. First, the Information Gap: standard skip connections pass spatial details from the encoder to the decoder indiscriminately. Since the encoder captures details for all objects in a scene while the decoder focuses on the salient one, this often introduces noise from non-salient background elements. Second, the Computational Cost of Sharp Edges: achieving crisp boundaries typically requires auxiliary edge-detection branches, which increase model complexity and inference time.
Proposed Method
To overcome these limitations, we introduce two novel components: the Context Fusion Decoder Network (CFDN) and the Near Edge Weighted Loss (NEWLoss).
Context Fusion Decoder Network (CFDN)
The CFDN treats global context as a filtering signal for skip connections. It consists of a Context Module (CM) that extracts a Global Salient Context Feature (GSCF) from the deepest encoder layer. This GSCF then guides the Feature Fusion Module (FFM), which suppresses non-salient spatial information from the encoder. The FFM utilizes parallel convolutional layers with varying dilation rates to capture multi-scale information effectively.
Near Edge Weighted Loss (NEWLoss)
To produce sharp boundaries without architectural overhead, we move edge-learning logic into the loss function. We generate a Near-Edge Mask (NEM) using distance transforms on the ground truth to highlight pixels near object boundaries. During training, the Binary Cross Entropy (BCE) loss is multiplied by this NEM weight map, forcing the model to prioritize accuracy on ambiguous edges. This approach allows the model to achieve the supervision effect of edge-reconstruction branches without any extra cost during inference.
Experimental Results
Our method was trained on the DUTS-TR dataset and evaluated across four benchmarks: ECSSD, PASCAL-S, DUT-OMRON, and DUTS-TE. Quantitative results show that our approach matches or exceeds state-of-the-art models like EGNet, MINet, and AFNet. Notably, our model achieved superior Structure Measure ($S_\alpha$) scores, indicating better preservation of global object structure.
Qualitative analysis confirms that the CFDN effectively ignores background noise, such as reflections, while the NEWLoss enables the segmentation of fine, complex structures like thin limbs or beaks. By using an EfficientNet-b5 backbone and avoiding dedicated edge branches, the model remains streamlined and efficient.
Conclusion
By rethinking feature fusion and boundary supervision, we demonstrate that SOD performance can be improved through smarter architectural integration and loss design. The CFDN and NEWLoss provide a streamlined, efficient path toward high-quality saliency mapping, bridging the gap between high-level semantic features and low-level spatial details.
Summary updated