Image Suitability Screening Using Rotated Chips for AI Models

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

AI/ML systems face challenges with bad imagery, such as blurry or noisy images, leading to erratic behavior, unreliable results, increased computational load, and reduced user trust due to inconsistent performance and increased false positives/negatives.

Innovation Solution

A system that assesses image quality by splitting images into chips, randomly rotating them, and using trained networks to estimate orientation accuracy, comparing it to actual orientation, and determining suitability based on a threshold, filtering out bad images.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If AI/ML systems process bad imagery (blurry, noisy, corrupt images), then the system can handle diverse input data, but the system exhibits erratic behavior, produces unreliable results, and increases computational load

Engineering Contradiction:

Improveability to handle diverse input dataVSAvoidsystem performance reliability

Core Design Contradiction:

Adaptability or versatilityVSReliability

Solution Approach 1:

The system performs preliminary quality assessment of images before they are processed by the AI/ML model. By evaluating image quality metrics (sharpness, noise level, corruption) in advance and filtering out bad imagery, the system ensures that only high-quality images are processed, preventing erratic behavior and unreliable results while maintaining the ability to handle diverse input data types

Inventive Principle:

Principle #10Preliminary action

2Productivity

If AI/ML systems process bad imagery, then the system can maintain high throughput, but false positives and false negatives increase

Engineering Contradiction:

Improveimage processing throughputVSAvoidtarget detection accuracy

Core Design Contradiction:

ProductivityVSMeasurement precision

Solution Approach 1:

The system performs preliminary quality assessment before processing images through the detection model. By filtering out bad imagery (blurry, noisy, corrupt images) in advance based on quality metrics, the system maintains high throughput for good images while preventing false positives and false negatives caused by processing low-quality images

Inventive Principle:

Principle #10Preliminary action

3Measurement precision

If the system filters out bad images through quality assessment, then image processing accuracy improves, but additional computational steps are required

Engineering Contradiction:

Improveimage quality assessment accuracyVSAvoidquality assessment process complexity

Core Design Contradiction:

Measurement precisionVSDevice complexity

Solution Approach 1:

The quality assessment process is segmented into independent evaluation modules that assess different image quality aspects (sharpness, noise level, corruption) separately. Each module processes specific quality attributes in parallel, reducing the computational burden while maintaining comprehensive assessment accuracy. This modular approach allows the system to filter out bad images effectively without excessive computational complexity

Inventive Principle:

Principle #1Segmentation

Data Source

PatentUS12536637B2Systems and methods for determining image suitability for trained models

Publication Date: 2026.01.27 COVAR LLC

AI summary

The present embodiments are generally related to determining the suitability of one or more images for artificial intelligence models responsible for image recognition. The systems include image processors, servers, and databased. The image processor can receive images, split them into image chips, randomly rotate the image chips, and grade an AI model's estimation of the image chip rotation compared against its true rotation.