Unsupervised anchor handling for machine vision system

A machine vision system comprising:

• a camera, wherein:
• the camera comprises an image sensor configured to capture images;
• and
• the camera is configured to use an anchor model, during an operating mode of the machine vision system, to derive a location of a candidate region of interest (ROI) relative to an anchor in a captured image;
• a manual input device;
• and
• a control device communicatively coupled to at least the camera and the input device, and comprising a processor configured, during a training mode of the machine vision system, to perform operations comprising:
• receive, from the camera, a set of training images captured by the image sensor, wherein each captured training image of the set of captured training images comprises multiple visual features and an ROI;
• receive, from the input device, manually input indications of a location of the ROI within each captured training image of the set of captured training images;
• perform at least one of a rotational transform, a translational transform, a scaling transform or a shear transform on the multiple visual features and ROI of at least one captured training image of the set of captured training images to at least partially align the multiple visual features and ROIs among all of the captured training images of the set of captured training images to identify a common set of visual features that are present within all of the captured training images;
• derive a converged ROI based on at least a portion of the ROI of at least one of the captured training images of the set of captured training images;
• and
• generate the anchor model based on a combination of the converged ROI and the common set of visual features, wherein:
• the common set of visual features defines the anchor that is different than the converged ROI;
• and
• a location of each visual feature is specified relative to the converged ROI.

A method comprising, during a training mode, performing operations comprising:

• receiving, at a processor and from an image sensor, a set of training images captured by the image sensor, wherein each captured training image of the set of captured training images comprises multiple visual features and a region of interest (ROI);
• receiving, at the processor, manually input indications of a location of the ROI within each captured training image of the set of captured training images;
• performing, by the processor, at least one of a rotational transform, a translational transform, a scaling transform or a shear transform on the multiple visual features and ROI of at least one captured training image of the set of captured training images to at least partially align the multiple visual features and ROIs among all of the captured training images of the set of captured training images to identify a common set of visual features that are present within all of the captured training images;
• deriving, by the processor, a converged ROI based on at least a portion of the ROI of at least one of the captured training images of the set of captured training images;
• and
• generating, by the processor, an anchor model based on a combination of the converged ROI and the common set of visual features, wherein:
• the common set of visual features defines an anchor that is different than the converged ROI;
• a location of each visual feature is specified relative to the converged ROI;
• and
• the anchor model is to be used by the processor during an operating mode to derive a location of a candidate ROI relative to the anchor in a captured image.