💻 computer science

Molten mark classification using multi-scale directional cross-scale attention fusion

This paper proposes a molten mark classification method for electrical fire investigation that leverages a Swin transformer and a bi-directional feature pyramid network with directional cross-scale attention fusion to achieve superior accuracy and robustness against image degradation compared to existing CNN-based approaches.

Original authors: Taehi Kim, Jonghwa Shim, Eenjun Hwang

Published 2026-08-03

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Taehi Kim, Jonghwa Shim, Eenjun Hwang

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a detective trying to solve a mystery, but the clues are tiny, melted blobs of metal left behind on a wire. In the world of electrical fire investigation, figuring out how a fire started is crucial. Was it a sudden, violent electrical short circuit that sparked a bead of molten metal? Or was it a slow, external heat source, like a candle or a heater, that just warmed the wire until it melted? The answer determines who is responsible and how to prevent it from happening again. Traditionally, experts had to cut these wires, polish them like precious gems, and look at them under expensive microscopes—a slow, expensive, and manual process. But recently, scientists have tried to teach computers to do this job by showing them pictures of the melted spots. The challenge is that these melted spots are tricky: they can look very similar, and the "clues" are hidden in both tiny details (like surface shine) and big shapes (like the overall boundary). It's like trying to tell two twins apart by looking at just one eye versus looking at their whole face; you need to see everything at once to get it right.

This paper introduces a new, super-smart computer brain designed specifically to solve this "melted metal mystery." The researchers built a system that acts like a team of detectives, each looking at the wire from a different distance. Some detectives zoom in super close to see the tiny scratches and shine (local details), while others step back to see the big picture and the overall shape (global context). The secret sauce of their invention is a special "attention" mechanism called Directional Cross-Scale Attention (DCSA). Think of this like a magical pair of glasses that doesn't just let the detectives see, but helps them talk to each other. If the "close-up" detective sees a shiny spot that looks like a short circuit, they can whisper to the "wide-angle" detective, "Hey, check if the whole shape matches a short circuit too!" This conversation happens in both directions, allowing the system to combine the best clues from every zoom level.

The paper finds that this new method is significantly better than the current best tools. When tested on a massive collection of over 18,000 images (including some taken in messy, real-life fire scenes), the new system achieved an F1 score of 0.9613 and a Matthews correlation coefficient (MCC) of 0.9257. To put that in perspective, it beat the previous top-performing model by 2.26 percentage points in F1 score and 4.02 percentage points in MCC. But the real magic happened when the images were messy. In the real world, fire scenes are often smoky, blurry, or poorly lit. When the researchers tested the models on "degraded" (blurry and noisy) images, the old computer models (based on standard CNNs) crashed, losing about 15.75% of their accuracy on average. The new system, however, only lost 2.88% of its accuracy. It stayed calm and accurate even when the clues were hard to see.

The authors also ran experiments to prove that their specific design choices were the reason for the success. They showed that simply adding standard "attention" (like a generic spotlight) wasn't enough; the system needed the specific "cross-scale" conversation where different zoom levels helped each other. They also proved that looking at the wire from both the top-down (global to local) and bottom-up (local to global) directions was essential; doing just one or the other made the system less accurate. By visualizing where the computer was "looking," they confirmed that their model focused exactly on the melted metal, ignoring distracting background noise like soot or grid lines, whereas older models often got confused by the background. Ultimately, the paper suggests that by letting different levels of detail talk to each other, we can build fire investigators that are not only faster and cheaper but also incredibly tough, capable of solving mysteries even when the evidence is a bit blurry.

Technical Summary: Molten Mark Classification Using Multi-Scale Directional Cross-Scale Attention Fusion

Problem Statement
In electrical fire investigations, distinguishing between short-circuit beads (caused by electric arcs) and heat marks (caused by external flames) is critical for determining fire origins and assigning liability. Traditional methods rely on metallographic analysis of wire cross-sections, which demands expensive equipment, significant manual labor for preprocessing (cutting, polishing), and time-consuming analysis. While recent deep learning approaches using Convolutional Neural Networks (CNNs) have shown promise, they face inherent limitations. CNNs utilize fixed-size convolutional kernels and local receptive fields, making them less effective at capturing the complex interplay of local textures (e.g., surface gloss) and global morphological features (e.g., boundary sharpness) required for accurate classification. Furthermore, real-world fire scene images often suffer from degradation (blur, noise, contamination), where fixed-receptive-field models struggle to maintain performance.

Methodology
The authors propose a novel classification architecture that integrates a Swin Transformer backbone with a modified Bi-directional Feature Pyramid Network (BiFPN) neck enhanced by a Directional Cross-Scale Attention (DCSA) module.

Backbone (Multi-Scale Extraction): The system employs a Swin Transformer to extract hierarchical visual features at four distinct resolution levels ( $C_1$ to $C_4$ ). This allows the model to capture fine-grained local textural details at high resolutions and global morphological context at lower resolutions. These features are projected to a common channel dimension ( $d=256$ ) via $1\times1$ convolutions.
Neck (DCSA-Enhanced BiFPN): Instead of using standard scalar-weighted summation in the BiFPN fusion nodes, the authors introduce the DCSA module. This module performs bidirectional feature fusion (top-down and bottom-up) while adaptively recalibrating features based on cross-scale dependencies:
- Top-Down Pathway: Propagates global context from coarse to fine scales. The DCSA module ( $DCSA_{td}$ ) applies channel attention (based on global average pooling of concatenated inputs) and spatial attention (based on cross-scale similarity and a spatial map generator) to the current-scale feature. A gated residual mechanism then adaptively fuses the attended feature with the upsampled coarse feature, controlling the propagation intensity of global context.
- Bottom-Up Pathway: Propagates local details from fine to coarse scales. The DCSA module ( $DCSA_{bu}$ ) processes three inputs: the downsampled bottom-up feature, the current-scale backbone feature, and the top-down intermediate feature. It employs a shared FC network for channel attention and a SoftMax-normalized spatial attention mechanism to select the most informative input at each spatial location, preventing redundant feature accumulation.
Classification Head: The fused features from all four scales are aggregated via Global Average Pooling (GAP), concatenated, and passed through a fusion network of fully connected layers (1024 $\to$ 512 $\to$ 256) with layer normalization, ReLU, and dropout, followed by a linear classifier for binary output.

Key Contributions

Novel Architecture: This is the first study to apply multi-scale feature fusion using a Swin Transformer and BiFPN specifically for molten mark classification.
DCSA Module: The authors propose the DCSA module, which replaces standard scalar weighting in BiFPN nodes with direction-aware channel and spatial attention. This allows the network to selectively emphasize classification-relevant channels and spatial regions conditioned on features from different resolution levels, addressing the limitation of existing BiFPNs that treat all channels and spatial locations equally.
Robust Dataset: The study enhances practicality by augmenting laboratory-generated samples with a dataset of molten marks collected from actual electrical fire scenes, which are often contaminated and degraded.
Comprehensive Validation: The method is validated through extensive comparative experiments, ablation studies, and qualitative analyses (GradCAM and t-SNE).

Experimental Results
The proposed method was evaluated on a dataset comprising 18,750 training images and 4,905 testing images (including both laboratory and real fire scene data).

Performance on Standard Dataset: The proposed model achieved an F1-score of 0.9613 and a Matthews Correlation Coefficient (MCC) of 0.9257. This represents an improvement of 2.26 and 4.02 percentage points, respectively, over the best-performing baseline (Swin Transformer). The model maintained balanced precision (0.9633) and recall (0.9593), outperforming CNN-based models (VGG16, EfficientNet) and other Transformer variants (ViT, DeiT).
Robustness to Degradation: On a degraded testing dataset (simulating blur, noise, and compression), the proposed model exhibited superior robustness, with only a 2.88% decrease in accuracy. In contrast, CNN-based models suffered average accuracy decreases of approximately 15.75%, often collapsing in precision due to their reliance on fine textures. The proposed model maintained an MCC of 0.8681, significantly outperforming baselines.
Ablation Studies: Removing either the top-down or bottom-up DCSA pathways, or isolating channel/spatial attention, resulted in performance drops, confirming that the full integration of cross-scale conditioning and bidirectional flow is essential for optimal performance.
Qualitative Analysis: GradCAM visualizations demonstrated that the proposed model focuses exclusively on molten mark regions, effectively suppressing background interference (e.g., gridlines, soot) where baseline models failed. t-SNE visualizations revealed that the model learns fine-grained sub-clusters within classes, differentiating diverse wire types and morphologies.

Significance
The paper claims that the proposed method addresses the critical limitations of existing CNN-based approaches by effectively leveraging both local and global features through multi-scale fusion. By introducing the DCSA module, the method selectively emphasizes discriminative cues that may be prominent at one scale but subtle at another. The results demonstrate that this approach not only achieves state-of-the-art accuracy but also provides the robustness necessary for real-world fire investigation scenarios where image quality is often compromised. The study suggests that this architecture offers a viable, automated alternative to labor-intensive metallographic analysis, potentially improving the efficiency and reliability of electrical fire cause determination.

More like this