Visual Object Appearance Model Weighting Outliers
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image recognition models in industrial inspection are prone to inaccuracies due to inclusion of outlier images during model creation, fail to provide reasons for non-conformity, and struggle to identify areas within images that best match the model, especially when those areas only partially satisfy the model.
Innovation Solution
A method for generating robust models by determining the influence of each image data item on the model based on its relationship with variability, using normalization techniques like median and median absolute deviation, and applying weights to minimize outlier impact, while also providing confidence values and probability distributions to assess image conformity and identify optimal areas.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional model generation methods are used that include all training images equally, then the model building process is simple, but the model accuracy deteriorates due to outlier images
Solution Approach 1:
The patent changes the parameter of data item influence from uniform weighting to variable weighting based on statistical analysis. Each data item is assigned a weight according to its relationship with the variability of the training set, transforming the model generation from a simple averaging process to a statistically robust process that accounts for outliers.
Solution Approach 2:
The patent performs preliminary statistical analysis (calculating variability, determining relationships between data items and variability) before generating the final model. This preliminary action identifies and downweights outlier images before they can negatively impact the model accuracy.
2Loss of information
If model determination only provides a binary conform/non-conform result, then the processing is simple, but diagnostic information is lost
Solution Approach 1:
The patent implements feedback by providing diagnostic information about which specific areas of an input image cause non-conformity. Instead of merely stating whether an image conforms to the model, the system identifies and highlights the problematic regions, giving operators actionable feedback for quality control.
Solution Approach 2:
The patent segments the diagnostic feedback by identifying specific areas or regions within the input image that best satisfy or violate the model. This segmentation allows operators to see exactly which parts of the image are problematic rather than receiving a generic non-conformity result.
3Reliability
If existing techniques are used to identify areas satisfying a model, then the process is straightforward, but areas that only partially satisfy the model are not identified
Solution Approach 1:
The patent applies partial action by identifying areas that partially satisfy the model rather than requiring complete satisfaction. This allows the system to highlight regions that meet some but not all model criteria, providing more nuanced and reliable area identification that reflects the actual degree of conformity.
Data Source
AI summary
A computer-implemented method of generating a model from a set of images. The method comprises processing a plurality of data items, each data item representing an image of said set of images, to determine variability between said plurality of data items; and generating model data representing said model based upon said data items and said variability, wherein the influence of each of said data items upon the generated model is determined by a relationship between a respective one of said data items and said variability.