Rule-Based Spatial Mixture-of-Experts U-Net for Explainable Edge Detection

1. Problem Statement

Edge detection is a foundational task in computer vision, yet a significant trade-off exists between accuracy and explainability:

• Classical Methods (e.g., Canny, Sobel): Offer complete transparency through mathematical definitions (gradient thresholds) but lack semantic understanding, often misclassifying texture noise as edges.
• Deep Learning Methods (e.g., U-Net, HED): Achieve state-of-the-art accuracy by learning hierarchical features but operate as "black boxes." Their decision-making processes are opaque, making it difficult to diagnose failures (e.g., hallucinated edges in smooth regions) or verify logic in safety-critical applications.

The authors aim to bridge this gap by creating an edge detection model that retains the high performance of deep learning while providing human-readable, rule-based explanations for its predictions.

2. Methodology: The sMoE U-Net

The proposed architecture is a hybrid neuro-fuzzy system that modifies the standard U-Net encoder-decoder structure with two key innovations:

A. Spatially-Adaptive Mixture-of-Experts (sMoE)

Integrated into the decoder's skip connections, the sMoE block replaces standard concatenation with a dynamic gating mechanism. It processes feature maps using two specialized "experts" based on local image statistics:

1. Context Expert (Smooth): Uses dilated 3×3 convolutions (dilation=2) to expand the receptive field. This captures regional context and suppresses texture noise in homogeneous areas.
2. Boundary Expert (Sharp): Uses pointwise 1×1 convolutions to focus strictly on local pixel intensity, preserving high-frequency details and ensuring crisp edge boundaries.
3. Gating Network: A learnable 1×1 convolution processes a heuristic edge signal (derived from Sobel filtering) to generate a soft mask $G(E) \in [0, 1]$. This mask dynamically weights the output of the two experts for every pixel:
   $$Y = G(E) \cdot \text{Expert}_{\text{sharp}}(X) + (1 - G(E)) \cdot \text{Expert}_{\text{smooth}}(X)$$

B. TSK Fuzzy Explainability Head

Instead of a standard sigmoid classification layer, the network terminates with a Takagi-Sugeno-Kang (TSK) Fuzzy Inference Engine.

• Inputs: The head takes a hybrid vector of Edge Strength (normalized Sobel magnitude) and Semantic Confidence (sigmoid activation from U-Net features).
• Logic: It learns $R=4$ first-order fuzzy rules. Each rule follows the form:
  $$\text{IF } x_1 \text{ is } A_{i,1} \text{ AND } x_2 \text{ is } A_{i,2} \text{ THEN } y_i = a_{i,0} + a_{i,1}x_1 + a_{i,2}x_2$$
• Output: The final edge map is a weighted average of the rule consequents, where weights are determined by the firing strength of the rules (product of Gaussian membership functions). This ensures the output is fully differentiable and interpretable.

3. Key Contributions

1. Dynamic Strategy Switching: The sMoE blocks explicitly model the trade-off between semantic suppression and boundary preservation, allowing the network to adapt its processing strategy (smoothing vs. sharpening) pixel-by-pixel.
2. Transparent Decision Logic: The TSK Fuzzy Head replaces opaque activation functions with explicit IF-THEN rules, enabling the system to articulate why an edge was detected (e.g., "High Gradient AND High Confidence").
3. Explainable Visualizations: The model generates two unique visualization tools:
   • Strategy Maps: Visualize which expert (Context or Boundary) is active in specific regions.
   • Rule Firing Maps: Visualize which logical rules are triggering for specific pixels, allowing engineers to trace decisions back to specific premises (e.g., distinguishing between a true edge and texture noise).

4. Experimental Results

The model was evaluated on the BSDS500 benchmark (200 training, 100 validation, 200 test images) using a composite loss function (Binary Cross Entropy + Dice Loss).

Quantitative Performance:

• ODS F-score: The sMoE U-Net achieved 0.7628, outperforming the standard U-Net (0.7437) and matching the specialized HED baseline (0.7688).
• Average Precision (AP): Achieved 0.7222, surpassing both U-Net (0.6946) and HED (0.7126), indicating superior stability across recall thresholds.
• Comparison to Classical Methods: Significantly outperformed Canny (0.5450) and Sobel (0.5769).

Qualitative & Explainability Findings:

• Visual Quality: Produced crisp, clean boundaries that aligned closely with Ground Truth, avoiding the "thick/noisy" edges of Sobel and the "hallucinated" edges of Canny in textured regions.
• Logic Verification:
  • Strategy Maps confirmed the network correctly activated the "Boundary Expert" for object outlines and the "Context Expert" for backgrounds.
  • Rule Firing Maps revealed distinct semantic behaviors:
    • Rule 1: Detected unambiguous edges (High Gradient + High Confidence).
    • Rule 2: Detected soft boundaries (Low Gradient + High Semantic Confidence).
    • Rule 4: Actively suppressed false positives (High Gradient + Low Confidence/Texture).

5. Significance and Impact

This work demonstrates that high performance and interpretability are not mutually exclusive in edge detection.

• Safety-Critical Applications: The "Glass-Box" nature of the model makes it suitable for domains where verification is mandatory, such as medical imaging (tumor boundaries) and aerospace maintenance (crack detection).
• Regulatory Compliance: The explicit rule-based logic allows domain experts to audit the model's decisions, satisfying regulatory requirements that black-box deep learning models often fail to meet.
• Future Direction: The authors propose extending this framework to higher-order TSK systems and validating it in real-world safety-critical scenarios.

In conclusion, the Rule-Based sMoE U-Net successfully bridges the gap between deep learning accuracy and fuzzy logic transparency, offering a new paradigm for auditable computer vision systems.