Visual Explanation for Image Analytics Decisions
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for explaining AI model predictions in image analytics lack the ability to visually indicate distinctive features, making it difficult for users to understand the reasoning behind AI decisions.
Innovation Solution
A method and system that uses deep learning to extract local features from images, aggregate them into a 2D pixel map, and superimpose this map onto the images to provide visual explanations by indicating mark-ups based on the degree of match, allowing for instance-level explanations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If global explanation methods are used to explain AI model predictions, then explanations can be provided for the overall model behavior, but they cannot provide instance-level detailed explanations for individual image predictions
Solution Approach 1:
The patent segments the explanation process into two distinct components: global explanation (using methods like anchors and feature importance for overall model behavior) and local explanation (using what-if analysis and counterfactuals for instance-level predictions). This segmentation allows the system to provide both broad model insights and specific prediction details simultaneously, resolving the contradiction between explanation scope and instance-level detail.
2Loss of information
If existing explanation methods are used that consider feature importance, then predictions can be explained based on individual pixels or pixel segments, but visual indications using mark-ups cannot be provided to easily explain image analytic decisions
Solution Approach 1:
The patent introduces visual mark-up elements as an intermediary between the AI model's feature importance calculations and the user's understanding. These mark-ups (such as highlighted regions, bounding boxes, or color-coded overlays) translate abstract feature importance data into intuitive visual representations, making image analytic decisions easily understandable while preserving all relevant feature information.
Data Source
AI summary
The invention relates to method and system for providing visual explanations for image analytics decisions. The method includes extracting a set of local features from each of a plurality of image instances using a deep learning (DL) model; determining a feature list by aggregating the set of local features from each of the plurality of image instances; generating a two-dimensional (2D) pixel map based on the feature list; superimposing the 2D pixel map of aggregated features on each of the plurality of image instances; and providing a visual explanation for an image analytics decision on one or more of the plurality of image instances based on superimposition.