Activation Map Artifact Mitigation for Reliable Saliency Maps
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Saliency maps in image classification models often contain artifacts that distort the actual relevance of highlighted image regions, impairing model interpretability and accuracy, particularly in automated optical inspection systems.
Innovation Solution
A method and device for optimizing activation maps by detecting and mitigating artifacts in feature maps using quantile-based deviation functions and transformation strategies to generate an optimized saliency map, reducing or eliminating artifacts.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of information
If saliency maps are generated using traditional methods (e.g., Grad-CAM), then model interpretability is provided, but artifacts appear in the saliency maps that distort actual relevance and impair interpretability
Solution Approach 1:
The patent applies preliminary action by detecting artifacts in activation maps before generating the saliency map. The method identifies artifact-affected regions in the feature maps using statistical tests (e.g., quantile-based outlier detection) and masks or corrects these regions prior to saliency computation, preventing artifact propagation to the final interpretability visualization.
Solution Approach 2:
The patent introduces an intermediary artifact detection and correction module between the neural network's feature maps and the saliency map generation process. This intermediary component analyzes channel-wise activation statistics, identifies outliers representing artifacts, and applies corrections (such as channel masking or value adjustment) before the corrected features are used to compute the final saliency map.
2Reliability
If artifact detection and correction steps are added to the saliency map generation process, then saliency map accuracy is improved, but computational complexity and processing time increase
Solution Approach 1:
The patent employs parameter changes by utilizing statistical parameters (quantiles, standard deviations, mean activations) of the feature map channels to detect artifacts. By changing the analysis from pixel-level to channel-level statistics, the method achieves artifact detection with computationally efficient operations that scale well with network depth and width.
Solution Approach 2:
The patent replaces complex iterative artifact removal mechanisms with direct statistical thresholding and channel-wise operations. Instead of using computationally intensive optimization or iterative refinement processes, the method applies closed-form statistical tests and simple masking operations that are much faster and less complex while achieving effective artifact suppression.
Data Source
AI summary
A method for optimizing activation maps, in particular for generating an optimized saliency map, includes (i) providing an activation or feature map which can be generated based on image data by a neural network, particularly a convolutional neural network, for each layer of the neural network and which indicates activation of channels of the neural network in response to the image data, (ii) at least for a plurality of pixels of the activation or feature map, determining pixel-by-pixel whether a quantile determinable across all channels is greater than a predetermined deviation function, or whether a magnitude across different quantiles is greater than a predetermined threshold value, in order to detect such an artifact the activation or feature map, and (iii) when an artifact is detected in the activation or feature map, applying a mitigation strategy to eliminate or at least reduce the detected artifact to provide an optimized activation or feature map.