Dual-Model Product Recognition for New Defect Detection

The patent divides the recognition system into two separate models: a defect recognition model and a qualification recognition model. Each model specializes in one aspect, allowing them to achieve high accuracy in their respective domains without the complexity of a single comprehensive model. The defect recognition model focuses on identifying defective products, while the qualification recognition model focuses on identifying qualified products, resolving the contradiction between accuracy and complexity.

The patent combines the outputs of two separate recognition models to make the final quality determination. By merging the defect recognition result and qualification recognition result, the system achieves comprehensive accuracy for both known and new defect types while maintaining the simplicity of individual specialized models.

The patent segments the recognition task into two specialized models with distinct functions. The defect recognition model handles known defect patterns, while the qualification recognition model handles overall product quality assessment. This segmentation allows each model to be optimized for its specific task, improving generalization capability while controlling computational resources through focused processing.

The patent applies partial action by using the qualification recognition model only when needed to resolve ambiguous cases or identify new defect types. For clear-cut cases, the system relies on the defect recognition model alone, thus reducing overall computational resource consumption while maintaining high adaptability for new defect scenarios.

The patent implements feedback mechanisms where the qualification recognition model's results are used to refine and retrain the defect recognition model. This feedback loop continuously improves model robustness by incorporating real-world performance data, allowing the system to adapt to new defect types while managing training time through iterative rather than exhaustive retraining cycles.

The patent employs periodic action by implementing feedback-based retraining at scheduled intervals or when performance degradation is detected, rather than continuous retraining. This approach enhances model robustness through regular updates while minimizing training time loss by avoiding constant retraining, thus resolving the contradiction between reliability and time efficiency.