Technical Summary: YOLO-LOLG Mamba for Industrial Steel Defect Detection

1. Problem Statement

Steel surface defects, such as cracks, inclusions, and scratches, are critical quality indicators in industrial manufacturing. However, existing target detection models, particularly those based on the YOLO (You Only Look Once) framework, struggle with two primary challenges in this domain:

• Loss of Small-Scale Features: Due to the small size of defects relative to the total image area, continuous convolution and down-sampling operations in standard architectures cause the loss of fine texture and edge details. While multi-scale fusion structures (e.g., FPN, PAN) exist, they often fail to retain sufficient shallow-level detail under lightweight constraints.
• Inadequate Modeling of Linear and Irregular Defects: Defects like scratches and cracks exhibit strong continuity and complex morphologies. Traditional convolutional layers are limited by local receptive fields, hindering the capture of long-distance dependencies and global context. Consequently, current models often miss edge continuity or fail to distinguish irregular defects from complex backgrounds.

2. Methodology: YOLO-LOLG Mamba

The authors propose YOLO-LOLG Mamba, a novel detection framework designed to enhance feature perception for small defects and model global context through cross-scale interaction. The architecture integrates three core modules into the YOLO structure:

A. Lightweight Omni-dimensional Dynamic Convolution (LODConv)

To address the loss of small defect features, LODConv is introduced to enhance shallow feature representation. It comprises:

• Lightweight Large Target Enhance (LTE) Module: Utilizes channel attention to enhance discriminative responses, followed by a 7×7 large-kernel spatial attention to expand the receptive field. This preserves defect boundaries and morphological information.
• Dual-Channel Attention & Adaptive Residual Fusion: These mechanisms improve the expression of shallow features while maintaining low computational complexity, providing richer local information for subsequent fusion.

B. Linear Defect Attention Module (LDAM)

Designed to tackle the modeling of linear, slender structures and fuzzy boundaries, LDAM builds upon the SimAM architecture with the following enhancements:

• Multi-Scale Branches: Integrates 3×3 and 5×5 convolution branches to extract local details and large-scale structural features, respectively.
• Local Contrast Calculation: Computes the difference between pixel values and local/global means to enhance the distinction between defect areas and backgrounds.
• Adaptive Fusion: Dynamically allocates weights across branches to collaboratively model global statistical information, local contrast, and edge features.

C. MsGroupMamba Layer

To achieve global context modeling and cross-scale feature fusion, the MsGroupMamba module is deployed in the "Neck" (feature fusion) layer rather than the backbone. This placement allows for global context interaction between multi-scale features with lower computational overhead.

• Lightweight Multi-Scale Fusion (LMSF): Located at the inlet, this module uses depthwise separable convolutions (1×1, 3×3, 5×5) combined with channel attention to extract local features and multi-scale contextual information.
• State Space Modeling: Integrates the global modeling capabilities of State Space Models (Mamba) with dual residual connections. This facilitates the collaborative fusion of local information and global context, enhancing the representation of both complex and small defects.

3. Key Contributions

The paper outlines three primary contributions:

1. Problem Identification: The study systematically identifies that existing YOLO-based methods lose small-scale features, struggle to model linear defects, and lack sufficient cross-scale interaction and global context modeling for steel defects.
2. Framework Proposal: The introduction of YOLO-LOLG Mamba, which combines LODConv and LDAM for enhanced local perception and MsGroupMamba for global context modeling, specifically tailored for industrial steel inspection.
3. Empirical Validation: Comprehensive experiments on the NEU-DET and GC10-DET datasets demonstrate significant improvements in detection accuracy and generalization compared to baseline YOLO models.

4. Experimental Results

The proposed method was evaluated on two datasets: NEU-DET (1,800 images, 6 defect types) and GC10-DET (2,300 images, 10 defect types).

• Performance on NEU-DET: YOLO-LOLG Mamba achieved an mAP@0.5 of 80.55%, representing a 3.94% improvement over the baseline YOLOv8n (76.61%). It also improved Precision to 80.34% and Recall to 74%.
• Efficiency: The model maintains a lightweight architecture with 2.36M parameters (approx. 31.2% fewer than YOLOv8) and 6.1 GFLOPs (13.4% reduction compared to YOLOv8), demonstrating a favorable balance between accuracy and inference speed.
• Generalization: On the more complex GC10-DET dataset, the model achieved an mAP@0.5 of 64.3%, outperforming YOLOv8 (63.7%) and other variants, verifying its robustness across different industrial scenarios.
• Ablation Studies:
  • Replacing standard convolutions with LODConv improved mAP by 1.04% while reducing parameters.
  • Replacing attention mechanisms with LDAM improved mAP by 1.44%.
  • The combination of all three modules (LODConv, LDAM, MsGroupMamba) yielded the highest performance, confirming their complementary nature.

5. Significance and Limitations

The authors claim that YOLO-LOLG Mamba successfully addresses the mechanistic limitations of standard YOLO frameworks in industrial settings by synergizing local detail enhancement, structural modeling, and global context modeling. The method achieves high detection accuracy without compromising computational efficiency, making it suitable for real-time industrial deployment.

However, the paper modestly notes its limitations:

• Dataset Scope: Validation was performed only on public datasets (NEU-DET and GC10-DET). Further evaluation is required for complex industrial environments involving illumination changes, noise interference, and class imbalance.
• Material Specificity: The model is currently designed for steel surface defects, and its cross-material generalization capability requires further investigation.
• Future Work: The authors suggest exploring cross-dataset transfer learning and domain adaptation to improve engineering practicability in broader scenarios.