Learning Crisp Edge Detector Using Logical Refinement Network

Luyan Liu, Kai Ma, Yefeng Zheng

Introduction

Edge detection is a fundamental task in computer vision, yet modern deep learning models often struggle with producing "crisp" results. Most state-of-the-art architectures generate thick, probabilistic boundaries that require extensive post-processing, such as Canny’s non-maximum suppression (NMS), to become usable. This problem is exacerbated in 3D medical imaging and biological research, where volumetric edge detection remains an under-explored and difficult frontier.

Researchers from Tencent Jarvis Lab have addressed these limitations by introducing the Logical Refinement Network (LRN). The core insight of this work is that an edge is, by definition, a subset of the object it bounds. By formalizing this mathematical relationship through a novel Logical Gate, the LRN produces high-quality binary edge maps directly from the network output, eliminating the need for traditional thinning algorithms.

Methodology

The LRN framework operates through a two-phase pipeline designed to transition from coarse global context to fine-grained local detail.

Phase 1: Joint Object and Edge Detection

In the first phase, the model simultaneously predicts a coarse object map ($D_O$) and a coarse edge map ($D_E$). This is achieved using an encoder-decoder architecture, similar to a U-Net, which is enhanced with Atrous Spatial Pyramid Pooling (ASPP). The ASPP module is critical for capturing multi-scale context, allowing the network to understand object boundaries at various resolutions.

Phase 2: Crisp Edge Refinement

The coarse maps generated in Phase 1 are fed into the refinement network. This stage focuses on sharpening the boundaries by leveraging the high-level object context to guide the low-level edge features.

The two-phase LRN framework architecture, showing the transition from joint detection to crisp refinement.

The Logical Gate Mechanism

The defining feature of the LRN is the Logical Gate ($G$). It enforces the constraint that edges must reside within the object area ($D_E \subset D_O$). The operation is defined as:

$$G = Conv(D_E \oplus (D_O \otimes D_E))$$

  • Multiplication ($\otimes$): The object map acts as a spatial mask, refining edge features and ensuring they do not bleed into the background.
  • Addition ($\oplus$): This step re-injects edge information to suppress noise and distractors within the segmentation features.

This gate is integrated into the skip connections of the refinement network, ensuring that every level of the decoder benefits from the logical relationship between objects and their boundaries.

Architecture of the Crisp Edge Refinement Network featuring the Logical Gate operation (inset) which uses object maps to mask and sharpen edge features.

Experimental Evaluation

The model was rigorously tested on two distinct and challenging datasets:

  1. 2D Nuclei Images: Utilizing 634 images from the Kaggle 2018 Data Science Bowl, characterized by dense, overlapping objects with thin boundaries.
  2. 3D Monkey Brain Microscopy: A volumetric dataset consisting of $300 \times 300 \times 100$ voxel patches acquired via light-sheet microscopy, testing the model's consistency across 3D space.

To handle the extreme class imbalance inherent in edge detection (where edge pixels are a tiny fraction of the total image), the authors employed a Hybrid Loss Function combining cross-entropy and focal loss.

Results and Discussion

The LRN demonstrated superior performance compared to established methods like HED, RCF, and BDCN:

  • Precision: On the 2D Nuclei dataset, LRN achieved a Dice Score (DSC) of 51.65%, outperforming BDCN (50.76%) and RCF (40.03%).
  • Localization: The model achieved the lowest Hausdorff Distance (HD) on the 3D Brain dataset (12.811 vs. BDCN's 18.341), proving that its predicted edges are physically closer to the ground truth.
  • Crispness: Qualitatively, the LRN produces sharp, single-pixel-width lines. Unlike other models that produce "halos" of uncertainty, the LRN output is binary-ready, tracking object boundaries with high fidelity.

Conclusion

The Logical Refinement Network represents a significant step forward in edge detection by treating the task as a logically constrained sub-task of segmentation. By integrating the Logical Gate, the model successfully bridges the gap between high-level semantics and low-level boundary features. This approach not only improves accuracy and localization but also streamlines the deployment pipeline by removing the need for heuristic post-processing steps.

Summary updated

BibTeX
@inproceedings{liu2020learning,
  author = {Luyan Liu and Kai Ma and Yefeng Zheng},
  title = {{Learning Crisp Edge Detector Using Logical Refinement Network}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2020},
  doi = {10.1007/978-3-030-59719-1\_33},
}
Figure