Semi-supervised method for training multiple pattern recognition and registration tool models

• US 9,679,224 B2
• Filed: 07/31/2013
• Issued: 06/13/2017
• Est. Priority Date: 06/28/2013
• Status: Active Grant

• Current Assignee: Cognex Corporation
• Original Assignee: Cognex Corporation
• Inventors: Michael, David J., Barker, Simon
• Primary Examiner(s): Liu, Li
• Application Number:
  US13/955,120
• Publication Number:
  US 20150003726A1
• Time in Patent Office: 1,413 Days
• CPC Class Codes: G06F 18/2155 characterised by the incorporation of unlabelled data, e.g. multiple instance learning [MIL], semi-supervised techniques using expectation-maximisation [EM] or naïve labelling
• CPC Class Codes: G06F 18/217 Validation; Performance evaluation; Active pattern learning techniques
• CPC Class Codes: G06F 18/2193 based on specific statistical tests
• CPC Class Codes: G06F 18/28 Determining representative reference patterns, e.g. by averaging or distorting; Generating dictionaries
• CPC Class Codes: G06V 10/772 Determining representative reference patterns, e.g. averaging or distorting patterns; Generating dictionaries

Abstract

A system and method for training multiple pattern recognition and registration models commences with a first pattern model. The model is trained from multiple images. Composite models can be used to improve robustness or model small differences in appearance of a target region. Composite models combine data from noisy training images showing instances of underlying patterns to build a single model. A pattern recognition and registration model is generated that spans the entire range of appearances of the target pattern in the set of training images. The set of pattern models can be implemented as either separate instances of pattern finding models or as a pattern multi-model. The underlying models can be standard pattern finding models or pattern finding composite models, or a combination of both.

Claims

1. A method for training a pattern recognition and registration model in a machine vision system, the method comprising the steps of:

providing one or more initial training images having a region specifying a pattern to be trained, the one or more training images being provided from a database containing a plurality of training images;

training a first pattern model from the one or more initial training images;

iterating over remaining images from the one or more initial training images and selecting a subset of high scoring images from the remaining images as input to model training; and

outputting a trained pattern model that includes features common to a predetermined number of the plurality of training images, the trained pattern model being different from the first pattern model.

2. The method as set forth in claim 1 wherein the step of iterating includes running the first pattern model to score each image.

3. The method as set forth in claim 1 wherein the first pattern model is trained using a first set of training parameters and a second pattern model is trained using a second set of training parameters.

4. The method as set forth in claim 1 wherein a metric used to score the remaining images comprises calculating a combined score which is initialized to zero, and if the pattern is found, using the first pattern model, in an image with a score greater than a user-defined confidence threshold, then that score is added to the combined score, and if the pattern is not found, using a first candidate pattern model, in an image with a score greater than the user-defined confidence threshold, then 1 is subtracted from the combined score.

5. The method as set forth in claim 1 wherein each feature in the trained output pattern occurs in approximately 80%-90% of the training images.

6. The method as set forth in claim 1 wherein the region specifying the pattern to be trained is given for each image by a predetermined ground truth.

7. The method set forth in claim 6 but where the predetermined ground truth is found for each image by running the first pattern model.

8. The method as set forth in claim 1 further comprising the step of training a second candidate pattern model having a second set of pattern training parameters, and iterating the second candidate pattern model over the remaining training images contained in the database and storing scores, poses and matching region data for the second candidate pattern model.

9. The method as set forth in claim 1 wherein the step of training the first pattern model further comprises storing scores, poses and matching region data.

10. The method as set forth in claim 1 wherein a first candidate pattern model comprises a composite model.

11. The method as set forth in claim 1 wherein the one or more training images that are provided from the database are selected by a computer.

12. The method as set forth in claim 1 wherein the trained pattern model is used in order to perform an alignment, a search or vision inspection tool in runtime operation of the machine vision system.

13. The method as set forth in claim 1 wherein a pattern origin is specified as input to training the first pattern model, in addition to the training image and region.

14. A system for generating pattern recognition and registration models, the system comprising:

a memory having computer executable instructions stored therein, the memory further comprising a database containing a plurality of training images, at least one image having a region specifying a pattern to be trained;

one or more processors that when executing the instructions are configured to;

train an initial pattern recognition and registration model by iterating the initial pattern recognition and registration model over the plurality of training images, and stores scores, poses and matching region data to provide a trained model; and

measure performance of the trained model over the plurality of training images.

15. A system for training a pattern recognition and registration model in a machine vision system, comprising:

a memory having computer executable instructions stored therein; and

one or more processors that when executing the instructions are configured to;

provide one or more initial training images having a region specifying a pattern to be trained, the one or more training images being provided from a database containing a plurality of training images;

train a first pattern model from the one or more initial training images;

iterate over remaining images from the one or more initial training images and selecting a subset of high scoring images from the remaining images as input to model training; and

output a trained pattern model that includes features common to a predetermined number of the plurality of training images, the trained pattern model being different from the first pattern model.