Scholarlink

<p>Edge detection constitutes a fundamental task in computer vision, serving as a pivotal component for applications such as image segmentation, object detection, and scene understanding. However, existing deep learning-based methods exhibit a critical yet often overlooked limitation: a reliance on pixel-wise loss functions (e.g., Binary Cross Entropy). By treating each pixel independently, these functions fail to supervise the intrinsic geometric properties of edges, such as continuity, smoothness, and directional consistency. This supervision gap is the root cause of common prediction artifacts, including fragmented lines, blurred boundaries, and structurally incoherent outputs. To address this challenge, we propose a novel Edge-Structure-Aware Loss Function as the core contribution of this study. Moving beyond traditional pixel-level matching, this function explicitly incorporates gradient constraints to enforce edge sharpness, continuity constraints to prevent fragmentation, and directional consistency constraints to ensure orientation coherence. Furthermore, we introduce a Three-Stage Dynamic Loss Scheduling Strategy, which leverages a curriculum learning process to guide the network from basic pixel-level learning to refined structural optimization. To provide a robust backbone for this structure-aware supervision, we design a dual-branch network architecture that integrates fine-grained semantic edge features with global contextual information via an efficient fusion module. Extensive experiments demonstrate that our method achieves highly competitive performance across multiple benchmarks, yielding ODS=0.847/OIS=0.861 on BSDS500, ODS=0.899/OIS=0.907 on Multicue, and ODS=0.761/OIS=0.776 on NYUDv2.</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

A CNN-transformer dual-branch network with structure-aware loss for high-resolution edge detection

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年3月19日

查看全文
  • Jinhao Jiang,
  • Junhao Guo,
  • Zijing Yang

摘要

Edge detection constitutes a fundamental task in computer vision, serving as a pivotal component for applications such as image segmentation, object detection, and scene understanding. However, existing deep learning-based methods exhibit a critical yet often overlooked limitation: a reliance on pixel-wise loss functions (e.g., Binary Cross Entropy). By treating each pixel independently, these functions fail to supervise the intrinsic geometric properties of edges, such as continuity, smoothness, and directional consistency. This supervision gap is the root cause of common prediction artifacts, including fragmented lines, blurred boundaries, and structurally incoherent outputs. To address this challenge, we propose a novel Edge-Structure-Aware Loss Function as the core contribution of this study. Moving beyond traditional pixel-level matching, this function explicitly incorporates gradient constraints to enforce edge sharpness, continuity constraints to prevent fragmentation, and directional consistency constraints to ensure orientation coherence. Furthermore, we introduce a Three-Stage Dynamic Loss Scheduling Strategy, which leverages a curriculum learning process to guide the network from basic pixel-level learning to refined structural optimization. To provide a robust backbone for this structure-aware supervision, we design a dual-branch network architecture that integrates fine-grained semantic edge features with global contextual information via an efficient fusion module. Extensive experiments demonstrate that our method achieves highly competitive performance across multiple benchmarks, yielding ODS=0.847/OIS=0.861 on BSDS500, ODS=0.899/OIS=0.907 on Multicue, and ODS=0.761/OIS=0.776 on NYUDv2.