💻 computer science

GLA-YOLO: A Lightweight YOLO-Based Network for Industrial Surface Defect Detection

This paper presents GLA-YOLO, a lightweight YOLO11-based network incorporating Lightweight Fusion Convolution and Dynamic Local Attention Fusion Blocks to achieve state-of-the-art real-time industrial surface defect detection with high accuracy and efficiency on resource-constrained edge devices.

Original authors: Jiayi Li, Norrima Binti Mokhtar, Wenjunliang Zhang, Muhammad Amirul Aiman Asri, Qiyuan Qin, Mahidzal Bin Dahari, Yixuan Dai, Bingxing Liu

Published 2026-10-30

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Jiayi Li, Norrima Binti Mokhtar, Wenjunliang Zhang, Muhammad Amirul Aiman Asri, Qiyuan Qin, Mahidzal Bin Dahari, Yixuan Dai, Bingxing Liu

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 quiet hum of a modern factory, the difference between a safe product and a dangerous one often comes down to a flaw invisible to the naked eye. A hairline crack in a steel beam, a tiny scuff on a car part, or a speck of corrosion on a casting can compromise the safety of an entire machine. For decades, factories have relied on human inspectors to scan these surfaces, a task that is slow, tiring, and prone to error when a worker's eyes grow heavy. To solve this, engineers have turned to computers, teaching them to see what humans miss. This field, known as computer vision, uses deep learning—a type of artificial intelligence that learns by studying thousands of examples—to spot defects automatically. However, teaching a computer to see tiny, faint flaws on a busy, textured background is incredibly difficult. The computer must be smart enough to find the needle in the haystack, yet fast enough to keep up with a conveyor belt moving at high speed, all while running on hardware that fits inside a factory control box.

A team of researchers from the University of Malaya and the Chinese Academy of Sciences has developed a new system called GLA-YOLO to tackle these exact problems. Their work focuses on creating a "lightweight" detection network, a type of software designed to be small and efficient enough to run on the edge of a factory floor without needing a massive supercomputer. The researchers started with a powerful existing framework known as YOLO11, which is already famous for its speed in finding objects. They realized that while this system was fast, it struggled with the specific quirks of industrial defects: the flaws often blend into the background, appear in wildly different sizes, and sometimes show up so rarely that the computer has almost no examples to learn from. To fix this, the team did not just tweak the existing code; they rebuilt key parts of the system from the ground up to be more sensitive to these difficult details.

The researchers introduced three main changes to the computer's "brain." First, they redesigned the initial layers that process the raw image. Instead of using a single, heavy method to look at the picture, they created a dual-pathway system. One path looks at the image in a standard way to catch general shapes, while the other uses a more specialized, efficient method to spot fine textures and edges. These two views are then combined, allowing the system to understand both the big picture and the tiny details without wasting energy. Second, they added a new module that helps the computer understand the relationship between distant parts of the image. In a standard system, looking at a long, winding crack might require the computer to compare every single pixel to every other pixel, a process that is slow and computationally expensive. The new system uses a streamlined approach that captures these long-range connections quickly, allowing it to see a defect that stretches across the entire surface as a single, coherent object rather than a collection of confusing dots. Finally, they integrated a smart gating mechanism that acts like a filter, telling the system to focus its attention strictly on the defect and ignore the noisy, irrelevant background patterns that usually confuse other detectors.

When the team tested their new system, the results were striking. They evaluated it on two different sets of data: a collection of 480 images of mechanical parts and casting blanks that they gathered from a real factory, and a public benchmark of steel surface defects containing ten different categories. On the factory images, which included rare defects like cracks that appeared only 19 times in the entire dataset, the new system achieved an accuracy of 90.21%. This was a significant improvement over the standard system it was based on, which scored 87.06%. On the more challenging steel benchmark, the new system reached 80.59% accuracy, again outperforming the standard version by a wide margin. Perhaps most importantly, the new system did this while being faster and lighter. It processed images at a rate of 168.57 frames per second, meaning it could analyze more than 168 images in a single second, a speed that is crucial for real-time factory lines. It also required fewer computational resources to run, using only 6.2 units of processing power compared to the 6.6 units needed by the standard system, despite being slightly more accurate.

The researchers also looked inside the system to understand why it worked so well. By visualizing where the computer was "looking" when it made a decision, they found that their new system focused sharply on the actual defects. In contrast, the older systems often got distracted by the texture of the metal or the lighting, spreading their attention across the background. The new system, however, learned to ignore these distractions, concentrating its energy on the cracks, scuffs, and corrosion spots. This ability to filter out noise was particularly evident in the rare defect categories, where the system managed to find the few existing examples without being confused by the lack of training data. The study confirms that by carefully redesigning how a computer processes visual information—making it smarter about where to look and how to connect distant details—it is possible to build a detector that is both highly accurate and fast enough for the demanding environment of modern manufacturing. This work suggests that the future of industrial quality control lies not in building bigger, more powerful computers, but in designing smarter, more efficient ways for them to see the world.

Technical Summary: GLA-YOLO for Industrial Surface Defect Detection

1. Problem Statement

Industrial surface defect detection is a critical component of quality assurance in manufacturing, yet it faces unique challenges that hinder the deployment of standard deep learning object detectors. These challenges include:

Low Contrast: Defects often exhibit extremely low contrast against textured or non-uniform industrial backgrounds.
Morphological Irregularity: Defects range from hairline cracks spanning a few pixels to extensive corrosion spalling, creating extreme scale variations.
Class Imbalance: Critical defect types may appear in fewer than twenty training samples (few-shot scenarios), leading to severe long-tailed distributions.
Resource Constraints: Factory environments require real-time inference on resource-constrained edge devices, a requirement that many high-accuracy architectures (e.g., those with quadratic attention complexity) fail to meet.

While the YOLO11 family offers a strong baseline for lightweight detection, the authors identify specific architectural limitations in the standard YOLO11n:

Shallow Feature Redundancy: Standard $3\times3$ convolutions in the early stem process all channels densely, incurring high computational overhead on high-resolution feature maps.
Over-parameterized Attention: The C2PSA module at the deepest backbone stage uses multi-head self-attention with quadratic complexity ( $O(N^2)$ ), which may be excessive for abstract semantic features.
Lack of Long-Range Context: Standard C3k2 blocks rely primarily on local convolutions, lacking explicit mechanisms to capture long-range spatial dependencies essential for detecting expansive or irregular defects.

2. Methodology: GLA-YOLO

The authors propose GLA-YOLO (Gated Local Attention YOLO), a lightweight architecture built upon the YOLO11 framework. It introduces three core innovations designed to enhance feature extraction and contextual modeling while maintaining a compact parameter budget.

2.1 Lightweight Fusion Convolution (LFConv)

Location: Replaces standard convolutional layers in the backbone stem (P1/2, P2/4) and neck downsampling paths.
Mechanism: LFConv employs a parallel dual-branch design:

Branch 1: A standard $3\times3$ convolution.
Branch 2: A depthwise separable convolution ( $3\times3$ depthwise + $1\times1$ pointwise).
Fusion: The outputs are concatenated, followed by Batch Normalization, SiLU activation, and a lightweight Channel Attention mechanism (SE-style) to adaptively recalibrate feature channels.
Goal: To enrich shallow feature representations with diverse extraction pathways while significantly reducing parameters and computational cost compared to dense convolutions.

2.2 Dynamic Local Attention Fusion Block (DLABlock)

Location: The fundamental computational unit integrated into composite modules.
Mechanism: DLABlock integrates four pathways:

Local Branch: A $3\times3$ depthwise convolution to preserve fine-grained spatial details and edge discontinuities.
Linear Attention Branch (DWConvAttention): Inspired by the Performer architecture, this module uses a kernel-based approximation (ReLU-based feature map $\phi(x) = \max(0, x) + \epsilon$ ) to decompose the attention mechanism. This reduces complexity from quadratic $O(N^2)$ to linear $O(N)$ , enabling efficient long-range context modeling without constructing the full $N \times N$ attention matrix.
Gated Fusion: A learnable gating mechanism (SE-style bottleneck) adaptively modulates the fusion of local and global features ( $F_{gated} = X_{local} + g \odot X_{attn}$ ), allowing the network to weigh local vs. global contributions dynamically.
CompactMLP: A post-fusion transformation module with an expansion ratio of 2 (reduced from the standard 4) to maintain parametric efficiency while providing non-linear representational capacity.

2.3 Composite Modules: DLAC3k2 and C2DLA

These modules embed DLABlock into the YOLO Cross Stage Partial (CSP) framework:

DLAC3k2: Replaces the standard C3k2 block at the deepest backbone stage (P5/32). It partitions features into two branches: one processed by DLABlock for attention-guided learning, and the other by standard Bottlenecks. This enables the capture of long-range dependencies in high-level semantic features.
C2DLA: Replaces the C2PSA module following the SPPF layer. It substitutes the standard multi-head self-attention with stacked DLABlocks, refining position-sensitive features with linear attention.

3. Key Contributions

Architectural Innovation: The introduction of LFConv, DLABlock, and their composite variants (DLAC3k2, C2DLA) specifically tailored for the constraints of industrial defect detection.
Efficiency-Accuracy Trade-off: The design achieves state-of-the-art performance among nano-scale detectors while reducing computational complexity (GFLOPs) and maintaining a parameter count comparable to YOLO11n.
Linear Attention Integration: Successfully adapting Performer-inspired linear attention for industrial defect detection to model long-range dependencies without the quadratic cost of standard transformers.
Few-Shot Capability: Demonstrated effectiveness in detecting rare defect categories (e.g., "CRACK" with only 19 instances) through improved feature localization and background suppression.

4. Experimental Results

The method was evaluated on two datasets:

Self-Collected Dataset: 480 images, 4 categories (Corrosion Spalling, Crack, Scuffing, Undercut), 1,367 instances.
GC10-DET Benchmark: 2,294 images, 10 steel surface defect categories.

Performance Metrics:

Self-Collected Dataset: GLA-YOLO achieved 90.21% mAP@50, outperforming the YOLO11n baseline (87.06%) by 3.15 percentage points. It also surpassed the significantly larger RT-DETR-18 model (20M parameters) by 2.74 points, despite having only 2.70M parameters.
GC10-DET: GLA-YOLO achieved 80.59% mAP@50, improving upon YOLO11n (76.19%) by 4.40 percentage points.
Efficiency: The model operates at 168.57 FPS (PyTorch) and 235.32 FPS (ONNX), which is faster than YOLO11n (160.79 FPS). It maintains a compact size of 2.70M parameters and 6.2 GFLOPs.

Ablation Studies:

DLAC3k2 provided the largest individual gain (+2.93% mAP@50 on GC10-DET), confirming the importance of long-range dependency modeling.
LFConv improved accuracy while simultaneously reducing parameters and GFLOPs.
C2DLA validated that linear attention can effectively replace multi-head self-attention with reduced cost.

Qualitative Analysis:
Class Activation Map (CAM) visualizations confirmed that GLA-YOLO learns spatially precise, defect-focused representations with reduced background attention, particularly in few-shot scenarios like the "CRACK" category.

5. Significance and Claims

The paper claims that GLA-YOLO establishes a new state-of-the-art performance among nano-scale detectors for industrial surface defect detection. Its significance lies in:

Solving the Edge Deployment Dilemma: It delivers high accuracy and real-time inference speeds on a compact model footprint, making it suitable for resource-constrained factory edge devices.
Addressing Specific Industrial Challenges: The architecture explicitly targets low-contrast defects, scale variations, and class imbalance through its gated fusion and linear attention mechanisms.
Pareto-Optimal Trade-off: The model achieves a simultaneous improvement in both accuracy and inference speed compared to the YOLO11n baseline, a rare achievement in deep learning optimization.

The authors note limitations, including the evaluation being restricted to the nano-scale and 640x640 resolution, and the self-collected dataset's limited size and scope (metallic surfaces). They suggest future work could explore rotated bounding boxes, semi-supervised learning for data scarcity, and deployment on ultra-low-power embedded platforms.