A computer vision model for industrial product surface inspection using unsupervised learning and few-shot learning

Abstract

The advancement of AI has significantly accelerated CV research in industrial manufacturing, focusing on using AI-based models to detect and classify surface defects in product inspection. However, existing models often underperform in practice due to neglecting defect characteristics. This study addresses three key issues: underutilization of large-scale normal samples, challenges in leveraging few defective samples, and ensuring robust detection against product/inspection standard changes. While industrial settings have abundant normal samples, effectively using them for training remains difficult. Humans can detect/classify defects with minimal examples, but deep learning requires substantial labeled data, which is often unfeasible. Lastly, current models struggle to generalize to new products/standards due to industrial products' vast diversity and irregularity. First, the unsupervised Reducing Biases (REB) is proposed for industrial anomaly detection representation. This model learns from normal images to detect product surface defects. A self-supervised task fine-tunes the pre-trained model via the DefectMaker strategy, ensuring diverse synthetic defects. Additionally, the local-density k-nearest neighbors (LDKNN) method addresses normal image patterns/distributions, reduces feature space density bias, and enables effective anomaly detection. Second, this research proposes a novel multi-view region context (MVREC) framework for few-shot defect multi-classification (FSDMC), focusing on generalization and contextual feature extraction. MVREC enhances defect classification by integrating pre-trained AlphaCLIP for general feature extraction and using a region-context framework with multi-view context augmentation. The framework also includes Few-shot Zip-Adapter(-F) classifiers to cache support set features for efficient few-shot classification. To validate MVREC, this study introduces MVTec-FS, a new FSDMC benchmark with instance-level mask annotations for 46 defect categories, and demonstrates its performance via comprehensive experiments on multiple datasets. Third, addressing fabric inspection challenges in the dynamic clothing industry--where new fabric types and inspection criteria emerge--this research proposes a Continual Learning-based Fabric Inspection Model (CLFIM). Traditional models fail to adapt to unseen patterns or defect categories due to pattern/criteria shifts. CLFIM, trained on specific fabrics, learns pattern and inspection criteria contexts to adapt to new tasks. Experiments on three complex-patterned fabrics with varying criteria show that the YOLOV8-based CLFIM effectively handles evolving inspection challenges. Overall, this research introduces innovative approaches--REB, MVREC, and CLFIM--to address critical challenges: underutilization of normal/defective samples and detection difficulties for altered products/inspection criteria. These models enhance defect detection/classification across fabric types and industrial products. Incorporating self-supervised, few-shot, and continual learning expands defect detection research scope.