Anomaly Detection for Small Defects in Optical Inspection

The patent segments the inspection task into two distinct processing paths: one for structural inspection (labeling, configuration) using traditional image analysis, and another for defect detection using deep learning anomaly detection. This segmentation allows each path to be optimized for its specific purpose, resolving the contradiction between precision for different inspection types and system versatility.

The patent introduces a new dimension to the inspection system by integrating deep learning-based anomaly detection alongside traditional optical inspection methods. This additional dimension enables the system to handle both structured inspection tasks (labeling, configuration) and unstructured defect detection tasks simultaneously, thereby improving both measurement precision and adaptability.

The patent segments the image processing into two resolution levels: high-resolution imaging for capturing detailed defect information, and then segmentation of processing tasks where anomaly detection operates on the high-resolution data. This allows detection of smallest anomalies while managing computational resources through targeted application of deep learning only where needed.

The patent applies local quality by using anomaly detection specifically for defect identification in regions where it is most needed, rather than uniformly processing entire high-resolution images. This localized application of computational resources improves detection accuracy for small anomalies while reducing overall computational burden.

The patent implements feedback mechanisms where the anomaly detection system continuously learns from inspection results and adjusts its detection parameters. This feedback loop ensures consistent quality assessment by maintaining objective detection criteria, eliminating the variability inherent in manual inspection while preserving operational simplicity through automated decision-making.

The inspection system performs self-service by automatically detecting and classifying defects without requiring human inspectors. The anomaly detection algorithm autonomously evaluates objects, ensuring consistent quality assessment while reducing reliance on human judgment, thereby improving reliability without sacrificing ease of operation.

The patent applies preliminary action by using anomaly detection to identify potential defects in non-defective objects during the training phase. This allows the system to proactively generate training examples of defects that would otherwise be rare in production data, thereby improving training data quality and volume balance before the model is deployed for actual inspection.

The patent converts the scarcity of defective samples (a harmful condition for training) into a benefit by using anomaly detection to synthetically generate defect training data from non-defective objects. This approach transforms the imbalance problem into an opportunity to create diverse training examples, improving both the quality and volume of training data available.