A Robust Industrial Defect Detection Method Integrating Capuchin Search and Sparse Representation

Technical Summary: A Robust Industrial Defect Detection Method Integrating Capuchin Search and Sparse Representation

Problem Statement
Automated visual inspection systems in smart manufacturing are increasingly vulnerable to sensor noise, illumination fluctuations, and adversarial perturbations, which compromise inspection reliability. While Extreme Learning Machines (ELMs) offer rapid training and favorable generalization for real-time industrial applications, their direct application faces two critical limitations:

1. Structural Denseness: Standard ELMs utilize fully connected architectures that incur high computational overhead and are prone to overfitting and sensitivity to irrelevant background clutter when processing high-dimensional industrial images.
2. Brittleness: ELMs, similar to deep neural networks, exhibit significant fragility when confronted with input perturbations, including imperceptible noise and adversarial attacks, rendering them unsuitable for safety-critical manufacturing environments.

Existing approaches often treat sparsity (for efficiency and generalization) and robustness (for noise tolerance) as orthogonal goals, lacking a unified framework to synergistically optimize both. Furthermore, manual hyperparameter tuning for complex sparse-robust architectures is inefficient and suboptimal.

Methodology: The CS-SR-IDD Framework
The paper proposes CS-SR-IDD (Capuchin Search with Sparse Representation for Industrial Defect Detection), a unified framework designed to address these limitations through three integrated components:

1. Sparse Feature Extraction (SPCA):
   The front-end employs Sparse Principal Component Analysis (SPCA) to process raw industrial surface images. Unlike standard PCA, SPCA incorporates an $L_1$ regularization penalty to enforce sparsity in the loading vectors. This achieves:

   • Dimensionality Reduction: Compressing high-dimensional pixel data into a compact subspace.
   • Denoising: Attenuating unstructured noise and disrupting dense inter-feature correlations that adversarial attacks exploit.
   • Interpretability: Generating sparse loadings where each principal component relies on a restricted subset of original pixels.

2. Robust Classification (Sparse ELM):
   The back-end utilizes a sparsified ELM classifier trained under selectable robust loss criteria (e.g., Huber loss, Maximum Correntropy Criterion, or Mean Absolute Error) rather than standard least squares. This allows the model to tolerate outliers and adversarial residuals that survive the SPCA filtering. The output weights are determined via an Iteratively Reweighted Least Squares (IRLS) paradigm.

3. Metaheuristic Optimization (Capuchin Search Algorithm - CSA):
   To eliminate the prohibitive cost of manual tuning, the framework employs the Capuchin Search Algorithm (CSA), a swarm-based metaheuristic simulating the foraging behaviors of capuchin monkeys. The CSA autonomously optimizes a multi-objective fitness function that balances:

   • Clean Accuracy ($A_{clean}$): Performance on unperturbed data.
   • Robust Accuracy ($A_{adv}$): Performance under adversarial perturbations.
   • Structural Parsimony ($S$): A penalty term discouraging overfitting based on model complexity.

   The algorithm simultaneously tunes the SPCA sparsity penalty ($\alpha$), the number of retained principal components ($k$), and the ELM hidden layer capacity ($h$), as well as the specific robust loss parameters.

Key Contributions

1. Unified Architecture: The proposal of a novel framework that integrates SPCA for disturbance-insensitive feature extraction with a robust ELM classifier, specifically tailored for industrial defect detection.
2. Automated Optimization Workflow: The design of a CSA-driven workflow that jointly tunes interdependent hyperparameters (sparsity constraints, subspace dimension, and network topology) to navigate high-dimensional search spaces effectively, outperforming manual calibration.
3. Comprehensive Evaluation Protocol: The establishment of a rigorous benchmarking protocol using CIFAR-10 and MNIST datasets (repurposed as proxies for industrial inspection) under diverse attack modalities (FGSM, PGD, MIM, SQUARE) and perturbation norms ($L_2$ and $L_\infty$).

Experimental Results
The author evaluated CS-SR-IDD against several baselines, including Raw-ELM, robust loss variants (ORELM, Huber-ELM, MCC-ELM), standard Back-Propagation (BP) networks, and non-optimized SPCA-ELM.

• Robustness: Under both $L_\infty$ and $L_2$ threat models, CS-SR-IDD consistently maintained higher accuracy than all baselines as perturbation intensity increased. Notably, under high-intensity iterative attacks (PGD, MIM), the model demonstrated a "smoother" performance decline, indicating more uniform decision boundaries.
• Ablation Studies: Experiments removing either the CSA optimizer (w/oCapuchin) or the SPCA module (w/oSPCA) confirmed that both components are critical. The SPCA module provided a foundational defense against structural perturbations, while the CSA optimization significantly amplified stability and adaptability, particularly in high-perturbation regimes.
• Generalization: The model demonstrated consistent robustness across different data distributions (CIFAR-10 vs. MNIST) and attack types, including black-box query-based attacks (SQUARE).

Significance and Claims
The paper claims that CS-SR-IDD provides an effective, highly optimized paradigm for constructing robust visual inspection systems in intelligent manufacturing. By synergistically coupling sparse representation with metaheuristic-driven hyperparameter optimization, the method achieves a favorable trade-off between classification fidelity and robust resilience against industrial interferences.

The author explicitly states that the current work serves as a validation of algorithmic efficacy using controlled benchmarks (CIFAR-10 and MNIST) as proxies. They acknowledge limitations, noting that robustness transferability to real-world industrial defects and physical noise sources requires further validation, and that the computational overhead of the CSA optimizer needs systematic cost-benefit analysis. Future work is directed toward real-world deployment on edge devices and testing against adaptive adversaries targeting the preprocessing stage.