Enhancing Small-Sample Defect Detection in Machinery Images via Multimodal Feature Perturbation

To address the core challenges in few-shot industrial defect detection—including sample scarcity, significant inter-class variations, and difficulties in micro-defect identification—this study proposes an innovative framework that integrates a Feature Perturbation Module (FPM) and a Cross-scale Pyramid Network (CSPN). Experiments conducted on a complex industrial dataset (containing five categories of mechanical components, such as gears, bearings, and transmission shafts, as well as six defect types) demonstrate the following: The proposed FPM module, through the synergistic optimization of geometric transformations and adversarial generation strategies, improves the geometric transformation coverage of crack samples by 40%, effectively suppressing overfitting. This results in an increase of 21.5 percentage points in F1-score compared to traditional CNNs. The CSPN network, leveraging a multi-level feature pyramid fusion mechanism, preserves high-resolution local features (128×128 pixels) while integrating deep semantic information. It achieves a recall rate of 82.3% for detecting sub-0.5mm micro-defects, surpassing the MAML baseline by 10.8 percentage points, and successfully identifies critical defects such as micro-cracks at gear edges, which conventional methods often miss. A weight optimization strategy (β²=0.3) reduces the false positive rate (FPR) to 4.2% in scenarios with oil contamination interference on bearings, improving detection accuracy by 12.7% compared to balanced-weight strategies. Ablation studies confirm module synergy effectiveness, showing significant advantages in F1-scores for three key components: gears (89.7%), bearings (85.2%), and shaft bushings (84.1%). The study further reveals performance bottleneck in extreme sample scarcity scenarios (≤5 samples per defect category), where model accuracy drops to 78.3%, and proposes an optimized approach combining dynamic weight adjustment and temporal feature analysis. This work provides a novel technical paradigm for advancing quality control systems in intelligent manufacturing.