AI Image Quality Assessment via Dropout Clustering
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Artificial intelligence (AI) systems that rely on image quality for detection and segmentation tasks become unreliable when image quality degrades due to environmental changes or other factors, leading to questionable detection and segmentation results.
Innovation Solution
A computer-implemented method that generates detection outputs for an image over multiple iterations by applying dropout randomly to different convolutional layers of a learning model, clusters the detection outputs, computes the total surface area for the clusters, and uses this as an uncertainty score to determine image confidence. If the confidence falls below a threshold, the system disables the detection system.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If the AI system operates on degraded images, then productivity is maintained, but reliability deteriorates
Solution Approach 1:
The system performs preliminary quality assessment of input images before processing. By evaluating image quality metrics (sharpness, noise level, lighting conditions) beforehand, the system can determine whether an image is suitable for AI processing, preventing unreliable detections from degraded images while maintaining continuous operation on quality images.
Solution Approach 2:
The patent introduces an intermediate quality assessment module between image input and AI processing. This intermediary evaluates image quality and either processes the image through the AI system or redirects it for capture/recapture based on quality thresholds, thus mediating between productivity requirements and reliability concerns.
2Measurement precision
If multiple iterations with dropout are applied, then measurement precision improves, but device complexity increases
Solution Approach 1:
The system applies dropout randomly to different convolutional layers across multiple iterations rather than processing the entire model exhaustively. This partial action approach provides sufficient precision for quality assessment while reducing overall computational complexity compared to full model processing.
Solution Approach 2:
The patent varies the dropout rate and iteration count as adjustable parameters to balance precision and complexity. By changing these parameters based on operational requirements, the system can achieve adequate measurement precision without consistently incurring maximum computational complexity.
Data Source
AI summary
Systems and methods include generating a detection output for an image over multiple iterations by applying a dropout randomly to a different convolutional layer of a learning model for each iteration. The detection outputs are clustered, on labels, for each iteration. A total surface area for the clusters is computed over the iteration. A confidence is computed for the image using the total surface area for the clusters as an uncertainty score. A system is disabled if the confidence is below a threshold.