TL;DR
This paper introduces VSS-ResNet, a novel dual-pathway deep learning architecture that combines the long-range dependency modeling of Mamba (a Vision State Space model) with the local feature extraction of ResNet for improved aero-engine ablation detection. This approach significantly enhances the accuracy and robustness of boundary segmentation for critical engine component defects, outperforming existing methods on a custom dataset.
What Problem Does This Paper Address?
Aero-engines operate under extreme conditions, leading to surface damage like ablation, which compromises structural integrity and flight safety. Detecting and quantifying this damage is crucial for predictive maintenance. Traditional methods, like manual visual inspection, are subjective and inefficient. While deep learning has shown promise, existing CNN and Transformer-based models struggle with the high computational cost and the precise boundary segmentation required for accurate ablation assessment, especially given the subtle, irregular, and low-contrast nature of these defects in borescope imagery.
Methods
The core innovation is the VSS-ResNet architecture, a dual-pathway network designed for semantic segmentation.
- Dual-Pathway Design: The network processes input images through two parallel branches:
- Left Path (VSS Path): Utilizes Vision State Space Blocks (VSSBs), inspired by Mamba. These blocks are designed to capture long-range dependencies in images with linear computational complexity (O(N)), making them efficient for processing large images. The VSSB itself incorporates layer normalization, depthwise separable convolutions, and a selective scanning mechanism (SS2D) for enhanced feature extraction.
- Right Path (ResNet Path): Employs a standard ResNet-50 backbone. This branch excels at extracting hierarchical local features and is adapted with atrous convolutions to maintain higher feature resolution.
- Multi-Scale Feature Fusion: Features from both pathways are progressively fused at intermediate stages (Stages 3 and 4) using concatenation and 1x1 convolutions. This fusion aims to combine the global context from the VSS path with the detailed local information from the ResNet path.
- PSPNet Decoder: The fused features are then fed into a Pyramid Scene Parsing Network (PSPNet) decoder for the final semantic segmentation, leveraging its pyramid pooling module for multi-level context aggregation.
Datasets
- Self-Constructed Engine Endoscopic Ablation Dataset: The primary dataset used for training and evaluation consists of 874 representative images acquired via borescope inspection of a specific aero-engine model. The dataset underwent a rigorous filtering process to remove redundant frames and ensure morphological diversity, followed by manual review by inspectors. Data augmentation techniques (rotation, cropping, flipping, histogram equalization) were applied to enhance diversity and model generalization.
Key Insights and Ideas
- Synergy of Global and Local Features: The paper strongly advocates for the necessity of combining long-range dependency modeling (Mamba/VSS) with local feature extraction (CNN/ResNet) for complex industrial defect segmentation. The dual-path architecture effectively leverages the complementary strengths of these paradigms.
- Efficiency of VSS for Long-Range Dependencies: Vision State Space models, particularly Mamba, offer a promising avenue for capturing global context in images with linear computational complexity, overcoming the quadratic bottleneck of traditional Transformers.
- Importance of Boundary Segmentation: The research highlights that accurate boundary detection is paramount for aero-engine maintenance, influencing severity classification and repair decisions. The proposed method's superior performance in boundary segmentation is a key contribution.
- Progressive Fusion for Richer Representations: Fusing features from the two branches at multiple scales allows for a more comprehensive representation that captures both coarse structural information and fine-grained textural details.
Key Results
- Superior Segmentation Performance: The VSS-ResNet achieved state-of-the-art results on their custom dataset, outperforming established models like U-Net, PSPNet, and DeepLabV3+ in terms of mIoU (81.71%), mPA (87.32%), and Acc (97.95%).
- Boundary Segmentation Improvement: Qualitative comparisons (Figure 10) show that the proposed method provides significantly more accurate and complete segmentation of ablation boundaries compared to other methods, which often exhibit over- or under-segmentation.
- Ablation Study Validation: Ablation studies confirmed that integrating VSSB into the left pathway significantly enhances performance across all metrics. Stacking multiple VSSBs on the same scale yielded diminishing returns, suggesting an optimal configuration. The dual-path architecture consistently outperformed single-path variants.
- Computational Trade-offs: While the dual-path VSS-ResNet incurs higher computational costs (parameters, FLOPs, inference time) compared to a baseline ResNet50, the performance gains are deemed worthwhile for the critical application of aero-engine inspection.
Limitations and Future Work
- Dataset Scale and Generalization: The authors acknowledge that the dataset, while carefully curated, is specific to one aero-engine model. Expanding the dataset scale and diversity is suggested to improve the model's generalization capability to different engine types and operating conditions.
- Computational Cost: The dual-path architecture, while effective, has a higher computational footprint than single-path models.
- Future Directions:
- Exploring integration with other advanced attention mechanisms or lightweight structures for further improvements.
- Developing more intelligent detection methods for comprehensive aero-engine health management.
Summary updated