Technical Summary: Multi-Scale Tensorial Summation and Dimensional Reduction Guided Neural Network for Edge Detection

1. Problem Statement

Edge detection is a fundamental low-level computer vision task critical for downstream applications such as image segmentation, enhancement, and intelligent transportation. Despite significant progress via deep learning, existing methods face three primary challenges:

1. Efficiency vs. Quality Trade-off: Achieving high qualitative results often requires deep, complex networks with large receptive fields, leading to high computational costs and redundant parameters.
2. Post-Processing Dependency: State-of-the-art (SOTA) methods frequently rely on non-maximum suppression (NMS) or morphological schemes to ensure result crispness, adding complexity to the inference pipeline.
3. Transfer Learning Reliance: Current top-performing models often depend on pre-training or transfer learning to handle challenging datasets like BSDS500 effectively.
4. Data Characteristics: Benchmark datasets suffer from label uncertainty (due to multiple annotations) and class imbalance (edges vs. non-edges), requiring robust handling of noise and redundancy.

2. Methodology: MTS-DR-Net

The authors propose MTS-DR-Net, a novel architecture designed to extract edge features efficiently by first removing redundant information rather than directly extracting necessary features. The network consists of two main components: an MTS-DR Backbone and a Refinement Network.

2.1. MTS-DR Backbone

The backbone is designed to identify and retain "necessary subspaces" (relevant edge information) while discarding homogeneous features.

• Multi-Scale Tensorial Summation (MTS) Layer: Inspired by Tucker decomposition and generalized tensor summation, the MTS layer replaces complex CNN or Transformer blocks. It operates on patch-wise feature maps of varying window sizes (e.g., 8, 16, 32, 64) to achieve large receptive fields from initial layers. It performs a learnable linear mapping via tensor summation:
  $$MTS(X) = \sum_{sc=1}^{SC} f^{-1}_{w_{sc}} \left( \sum_{t=1}^{T} f_{w_{sc}}(X) \times_1 A^{(t,sc)}_1 \times_2 \dots \times_J A^{(t,sc)}_J \right)$$
• Multi-Head Gate (MHG) Layer: To introduce non-linearity without traditional activation functions (like ReLU), the MHG layer uses a gated operation. It splits feature maps into heads, applying learned linear maps ($g_1, g_2$) via $1\times1$ and $3\times3$ convolutions, followed by element-wise multiplication. This approximates the curvature of the signal space.
• Dimensional Reduction Mechanism: An MTS-DR block sequentially applies an MTS layer (compressing dimensions), an MHG layer (non-linearity), and a second MTS layer (restoring dimensions). The final output of the backbone is a feature map $\tilde{X}$ generated by subtracting the sigmoid-gated input from the original input:
  $$\tilde{X} = X - \text{Sigmoid}(X) \otimes X$$
  This operation effectively suppresses redundant features, highlighting discontinuities.

2.2. Refinement Network

The backbone output feeds into a lightweight, U-shaped refinement network.

• Architecture: It utilizes an encoder-decoder structure with three scales. The encoder downsamples features, while the decoder upsamples them.
• Side Fusion: Features from three scales are fused using a Conditional Convolutional (CondConv) layer with four experts to enhance learning capacity at each scale.
• Output: The network produces three side outputs and a final fused edge map via a $1\times1$ convolution and Sigmoid activation, generating probability maps ranging from 0 to 1.

2.3. Loss Function

To address label uncertainty and class imbalance, the model employs a Weighted Binary Cross-Entropy (WBCE) loss.

• It balances positive (edge) and negative (non-edge) pixels using weights $\alpha$ and $\beta$.
• A threshold $\eta$ is introduced to mitigate uncertainty in ground-truth annotations (ignoring pixels with ambiguous labels).
• The total loss sums the WBCE of the three side outputs and the final fused output.

3. Key Contributions

1. Pioneering Tensorial Summation: This work is the first to utilize a tensorial summation strategy specifically for edge detection tasks, serving as a compact substitute for complex deep structures.
2. Redundancy-First Approach: Unlike conventional methods that extract features directly, the proposed model first removes unnecessary features via dimensional reduction, focusing the network on relevant subspaces.
3. Efficiency Without Post-Processing: The MTS-DR-Net achieves high precision and efficiency simultaneously without requiring NMS or other post-processing steps.
4. Performance Without Transfer Learning: The model surpasses previous SOTA methods on benchmark datasets even without the use of transfer learning or pre-training.

4. Experimental Results

The model was evaluated on BSDS500 and BIPEDv2 datasets, comparing against TEED, PiDiNet, DexiNed, and XYW-Net.

• BSDS500: The MTS-DR-2 variant achieved the best Optimal Dataset Scale (ODS) of 0.744 and Optimal Image Scale (OIS) of 0.758 (Thin setting), outperforming XYW-Net by 1.9% in ODS. The lightweight MTS-DR-3 variant achieved comparable performance to XYW-Net with 40.9% fewer parameters and 56.9% lower GFLOPs.
• BIPEDv2: The model achieved a mean IOU of 0.6038 and mean Precision of 0.8689. The MTS-DR-2 variant achieved the best ODS (0.903) and OIS (0.909).
• Ablation Studies:
  • Scalability: Increasing the number of MTS-DR blocks ($M$) and channels ($C$) improved performance, but smaller configurations (MTS-DR-3) remained highly competitive.
  • Compression Ratio (CR): A lower compression ratio (0.2) yielded slightly better ODS/OIS scores with reduced computational complexity compared to 0.4.
• Qualitative Analysis: Visual results indicate that the MTS-DR backbone successfully removes homogeneous features, resulting in edge maps with richer details and less noise compared to competitors.

5. Significance and Claims

The paper claims that MTS-DR-Net represents a significant shift in edge detection architecture by leveraging multi-scale tensorial summation to achieve large receptive fields early in the network. Its primary significance lies in:

• Eliminating Redundancy: By focusing on dimensional reduction to isolate necessary subspaces, the network avoids the computational burden of processing irrelevant features.
• Simplicity and Robustness: The method achieves SOTA performance without the need for complex pre-training pipelines or post-processing morphological operations.
• Efficiency: It demonstrates that high-quality edge detection can be achieved with significantly fewer parameters and lower computational costs (GFLOPs) than existing deep learning approaches, making it suitable for resource-constrained environments.

The authors conclude that the multi-scale tensorial summation factorization operator is a viable and efficient replacement for complex DNN structures in edge detection tasks.