Efficient Deep Learning Model Adaptation
MVTec Software introduces a new deep learning feature in Halcon 25.11, enabling flexible adaptation to evolving production environments
Continual learning enhances the retraining of deep learning models, focusing on efficiency and speed, and requires only a few images, reducing effort and costs. This feature addresses the need for retraining when production parameters change, such as lighting conditions or supplier variations, which can affect quality control and defect detection. Continual learning allows existing classification models to be quickly adapted to new requirements, typically needing only five to ten images.
It prevents "catastrophic forgetting," ensuring the neural network retains previously learned classes and features during retraining. This feature operates on a standard CPU, making it suitable for embedded and edge environments like smart cameras and sensors. It also supports the addition of new classes without full model retraining, allowing users without deep image processing expertise to perform retraining effectively.
