💻 computer science

LAC-Net: A Lightweight Attention and Category-Aware Reweighting Network for Steel Surface Defect Detection

LAC-Net enhances steel surface defect detection by integrating a lightweight attention module and a category-aware reweighting strategy into YOLOv11, achieving superior performance on the long-tailed GC10-DET dataset through improved handling of faint defects and class imbalance, though its effectiveness varies depending on the specific data distribution.

Original authors: Jianyong Zhao, Zigao Wu

Published 2026-10-08

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Jianyong Zhao, Zigao Wu

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

In the vast, humming factories where steel is rolled into sheets and coils, a single missed flaw can cost a company thousands of dollars. A scratch that is too small to see with the naked eye, or a patch of rust that blends into the background, can turn a perfect roll of metal into a rejected product. For decades, human inspectors have scanned these surfaces, but their eyes tire, and their attention wanders. To solve this, engineers have turned to computer vision, teaching machines to see what humans miss. The most common tools for this job are object detectors, software that scans an image and draws boxes around defects. These systems work well when the problems are obvious and the data is balanced, but steel factories present a unique challenge: the defects are often tiny, covering only a few dozen pixels, and some types of flaws appear far more often than others. When a computer learns from such uneven data, it tends to ignore the rare problems, focusing only on the common ones it sees every day.

Researchers at North China Electric Power University have developed a new system called LAC-Net to fix this specific weakness. Built on a modern detection framework known as YOLOv11, their approach does not try to be a massive, complex supercomputer. Instead, it is designed to be lightweight and efficient, making it suitable for the real-world constraints of a factory floor. The team made two key changes to how the software thinks. First, they tweaked the part of the system that looks at the big picture. They replaced standard processing steps with a new, low-cost method that helps the computer pay attention to the most important details, like the faint edges of a scratch or the subtle texture of a defect, while ignoring the noisy background. Second, they changed how the computer learns from its mistakes. Because some defects are rare, the system was taught to give extra weight to those difficult examples during training, ensuring it does not simply ignore them in favor of the common ones.

The results of this work were tested on two different collections of steel defect images. On the first dataset, which features a wide variety of flaws with very different frequencies, the new system showed a clear improvement. It detected more defects and located them more precisely than the standard version of the software, all while using fewer computer resources. The researchers found that the biggest boost in performance came from the strategy of reweighting the rare classes, proving that teaching the computer to care about the uncommon defects was the key to success. However, the story did not end with a universal victory. When the researchers tested the exact same system on a second dataset with a different mix of defects, the results were more mixed. While the system still performed well, the specific tricks that worked so well on the first set of data did not provide the same advantage on the second.

This outcome revealed a crucial insight about industrial inspection: there is no single "perfect" setting that works for every factory. The strategy that helps a computer learn from a long list of rare defects in one environment does not automatically transfer to a different environment where the defects are distributed differently. The researchers concluded that while the new attention mechanism helps the computer see better in general, the training strategy must be adapted to the specific data at hand. The work demonstrates that the future of automated inspection lies not in finding a single magic solution, but in building flexible systems that can be tuned to the unique rhythm of each production line. By combining a sharper way of seeing with a smarter way of learning from rare examples, LAC-Net offers a practical path forward for catching the tiny flaws that matter most.

Problem Statement

Steel surface inspection faces unique challenges that hinder the performance of standard deep learning detectors, particularly those based on the YOLO architecture. The primary difficulties include:

Small and Subtle Defects: Defects often occupy only a few dozen pixels with weak texture contrast, making localization and classification difficult.
High Inter-class Similarity: Different defect types (e.g., inclusions vs. scratches) share nearly identical textures, differing mainly in location, leading to ambiguous decision boundaries.
Long-Tailed Class Imbalance: Industrial datasets, such as GC10-DET, exhibit severe class imbalance where frequent categories dominate training, causing the model to underweight rare classes and faint defect regions.
Generalization Limits: Strategies optimized for one dataset often fail to transfer effectively to others with different distribution characteristics.

Existing detectors often underweight faint defect regions in high-level features, and their objectives are dominated by head classes, resulting in missed rare defects.

Methodology: LAC-Net

The authors propose LAC-Net, a detector built upon the YOLOv11 baseline. The architecture follows the standard Backbone-Neck-Head paradigm but introduces two core innovations to address the specific challenges of steel defect detection:

1. Lightweight Attention Enhancement (LAE) Module

Inserted into the high-level semantic stages of the backbone, the LAE module replaces standard convolutions to improve feature representation without significantly increasing computational cost.

Mechanism: It combines a low-cost feature transformation (replacing standard convolutions with efficient linear operations) followed by channel recalibration based on the Squeeze-and-Excitation (SE) scheme.
Function: The module applies Global Average Pooling (GAP) to an intermediate representation, processes it through fully connected layers with a reduction ratio, and generates a scaling vector. This vector reweights the input feature map, emphasizing channels relevant to defect recognition while suppressing redundant background responses.
Efficiency: This design reduces the parameter count and computational cost (GFLOPs) compared to the baseline while enhancing sensitivity to fine-grained regions and blurred boundaries.

2. Category-Aware Reweighting Strategy

To address class imbalance, the authors propose a multi-faceted reweighting strategy that acts on three distinct training components:

Bounding-Box Regression: Utilizes Wise-IoU (WIoU) as the localization loss. It assigns a dynamic weight to each sample based on its regression quality, down-weighting both high-quality and low-quality outliers to focus training on ordinary-quality samples, thereby improving boundary prediction.
Sample Selection: Implements a balanced sampling mechanism during data loading. The probability of sampling a category is inversely proportional to its frequency (controlled by a balancing strength $\gamma$ ), ensuring minority classes appear more frequently in training batches.
Classification Supervision: Applies adaptive class weights to the classification loss. The supervision intensity for tail classes is dynamically increased based on their frequency, ensuring they receive stronger learning signals.

3. Strategy Adaptation

The paper adopts a unified codebase where the model architecture remains fixed, but training strategies (loss functions, sampling, weights) are switched based on the target dataset's characteristics.

Key Contributions

Architectural Innovation: A lightweight attention enhancement module that replaces standard convolutions in high-level backbone stages, improving mAP while reducing parameters and GFLOPs.
Holistic Reweighting: A category-aware strategy that simultaneously addresses class imbalance in regression, sampling, and classification branches.
Empirical Boundary Definition: A quantified analysis demonstrating that while the full configuration improves performance on long-tailed datasets, it is not universally transferable. The effectiveness of the reweighting strategy is shown to be highly dependent on the specific class distribution of the target dataset.

Experimental Results

Performance on GC10-DET (Primary Dataset)

On the GC10-DET dataset, which features significant class imbalance:

Accuracy: LAC-Net improved mAP@50 from 64.37% (YOLOv11n baseline) to 66.51% (+2.14 points) and mAP@50:95 from 33.58% to 35.48% (+1.90 points).
Efficiency: The model reduced parameters from 2.59 M to 2.39 M and GFLOPs from 6.45 to 6.12.
Ablation Insights: The ablation study revealed that Category Reweighting (CR) alone was the most significant contributor, raising mAP@50 to 65.97%. The full combination of LAE, WIoU, Balanced Sampling (BS), and CR yielded the best results. Notably, using WIoU or BS in isolation without structural enhancement actually performed below the baseline, indicating that training strategies and architecture must be tuned jointly.

Generalization on NEU-DET (Auxiliary Dataset)

On the NEU-DET dataset, which has a more balanced distribution:

Performance: The full LAC-Net configuration achieved 76.81% mAP@50 (slightly above the baseline's 75.06%) but 42.75% mAP@50:95 (slightly below the baseline's 42.80%).
Observation: The gains observed on GC10-DET did not fully transfer. The baseline YOLOv11n remained highly competitive, suggesting that the specific reweighting strategies optimized for long-tailed distributions are not universally beneficial and can even slightly degrade performance on balanced datasets.

Significance and Claims

The paper claims that LAC-Net offers a practical solution for industrial steel surface inspection by balancing accuracy and efficiency. Its significance lies in two main areas:

Data-Aware Optimization: It demonstrates that a "one-size-fits-all" approach to training strategies is insufficient. The best module combination is dataset-dependent. While lightweight attention improves feature representation, the category-aware reweighting strategy is most effective when the target data exhibits a long-tailed distribution.
Engineering Practicality: The framework supports a unified codebase where the model structure is shared, but optimization strategies (loss, sampling, weights) are adapted flexibly to the target dataset. This allows for efficient deployment across different industrial scenarios without requiring entirely new model architectures.

The authors conclude that future work should focus on designing lightweight enhancement modules with stronger cross-dataset robustness and developing adaptive strategy selection mechanisms based on data-distribution statistics to further improve generalization.