Scholarlink

<p>Salient object detection (SOD) remains a pivotal yet challenging task in computer vision, particularly for high-resolution images where existing methods face two critical limitations: (1) The recognition and segmentation of target edges and semantic structures with significant variations becomes challenging due to limited cross-region adaptability capability in high-resolution images. (2) Low-contrast regions present in high-resolution complex scenes can lead to feature space representation confusion due to semantic ambiguity, which in turn triggers target-background classification ambiguity. To address these limitations, this paper proposes a multi-feature interactive guidance network (MIGNet) for high-resolution SOD. The framework employs an interactive guidance network as the backbone for saliency inference, augmented with three specialized components. First, an adaptive multi-scale feature enhancer dynamically regulates cross-region feature responses to prevent larger targets from dominating while maintaining sensitivity to fine structures. Second, a probabilistic context aggregation module tackles semantic uncertainty through probability-driven feature disentanglement and local-global probability correction, effectively resolving classification ambiguity in low-contrast regions. Third, a cross-scale attention fusion module combines upsampling-guided feature reorganization with bidirectional attention mechanisms to eliminate redundant responses while preserving precise spatial-semantic relationships for accurate edge delineation. The complete architecture is further optimized through a carefully designed loss function that progressively refines saliency predictions. Extensive experimental evaluations across nine benchmark datasets demonstrate that our MIGNet delivers state-of-the-art performance across diverse detection scenarios, consistently achieving superior results for large targets, clustered multi-targets, and isolated simple targets under varying complex conditions. The framework achieves remarkable MAE scores of 0.034, 0.032, 0.028, 0.050, 0.056, 0.090, 0.015, 0.030, and 0.026 on nine databases, while maintaining exceptional computational efficiency with merely 28.25M parameters and 70.63G FLOPs computational overhead.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Multi-feature interactive guidance network for high-resolution salient object detection

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月1日

查看全文
  • Baoyu Wang,
  • Mao Yang,
  • Pingping Cao

摘要

Salient object detection (SOD) remains a pivotal yet challenging task in computer vision, particularly for high-resolution images where existing methods face two critical limitations: (1) The recognition and segmentation of target edges and semantic structures with significant variations becomes challenging due to limited cross-region adaptability capability in high-resolution images. (2) Low-contrast regions present in high-resolution complex scenes can lead to feature space representation confusion due to semantic ambiguity, which in turn triggers target-background classification ambiguity. To address these limitations, this paper proposes a multi-feature interactive guidance network (MIGNet) for high-resolution SOD. The framework employs an interactive guidance network as the backbone for saliency inference, augmented with three specialized components. First, an adaptive multi-scale feature enhancer dynamically regulates cross-region feature responses to prevent larger targets from dominating while maintaining sensitivity to fine structures. Second, a probabilistic context aggregation module tackles semantic uncertainty through probability-driven feature disentanglement and local-global probability correction, effectively resolving classification ambiguity in low-contrast regions. Third, a cross-scale attention fusion module combines upsampling-guided feature reorganization with bidirectional attention mechanisms to eliminate redundant responses while preserving precise spatial-semantic relationships for accurate edge delineation. The complete architecture is further optimized through a carefully designed loss function that progressively refines saliency predictions. Extensive experimental evaluations across nine benchmark datasets demonstrate that our MIGNet delivers state-of-the-art performance across diverse detection scenarios, consistently achieving superior results for large targets, clustered multi-targets, and isolated simple targets under varying complex conditions. The framework achieves remarkable MAE scores of 0.034, 0.032, 0.028, 0.050, 0.056, 0.090, 0.015, 0.030, and 0.026 on nine databases, while maintaining exceptional computational efficiency with merely 28.25M parameters and 70.63G FLOPs computational overhead.