Image inspection identifies objects of interest such as defects. The appearance of objects of interest is often not well defined. They can have any size, shape and contrast. Machine vision systems detect and characterize objects of interest, apply decision rules to discriminate between true objects of interest (in this example - defects) and normal objects (false alarms), and classify true objects of interest into different classes or grades.
Decision rules encapsulate the knowledge acquired from the application and transform it into recipes to make decisions on new data. Decision rules are responsive to the training used to create them. However, they do not necessarily yield robust performance in the application they were intended to service. General decision rules that are domain independent generally do not perform well. Yet, expert systems that are highly domain specific are frequently not robust.
DRVision Technologies LLC decision learning technology automates the process of knowledge acquisition rather than relying on artisans to develop and debug esoteric domain specific knowledge bases. DRVision Technologies LLC decision learning separates noise and consistent characteristics from the application data and uses consistent characteristics to create robust decision rules that work well for the application.
The technology used to improve robustness moves a step beyond traditional decision rule structures. DRVision Technologies LLC decision learning technology regulates the decision rules to respond appropriately to uncertainty.It automatically adjusts the operating characteristic between crisp and soft decisions to match the application. It provides automatic optimization of decision rules by assessing the robustness and generalization of the decisions.
After the initial design, things sometimes change. DRVision Technologies LLC decision learning technology allows incremental update of decision rules and structures using newly acquired samples from existing or new classes. DRVision Technologies LLC technology offers the advantage of graceful incremental update; that is, the characteristics of the decision rules change smoothly and predictably in the incremental update process. Using new samples from existing classes, classifier statistics are incrementally updated to gracefully optimize performance for new samples. When new samples are added from new classes, a compound classifier structure is constructed to handle the new classes yet maintain stable performance for samples from existing classes. Previous experience is not lost. “change” is not the same as “start over”.
DRVision Technologies LLC decision learning technology produces human comprehensible rules that allow users to study and audit the decision rules. This assures the quality of the decision and offers possibilities for information integration and data mining applications.