Saliency Analysis Using Low- and High-Order Attention Features

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing saliency evaluation methods, both bottom-up and top-down, fail to accurately predict line-of-sight guidance in information processing and provide understandable improvement directions for design enhancements.

Innovation Solution

A saliency analysis system that combines low-order and high-order image feature amounts to generate saliency maps, incorporating human psychological tendencies and image features, allowing for precise saliency determination and actionable improvement suggestions.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If deep learning is used for line-of-sight prediction, then prediction accuracy is improved, but interpretability of which factors guide attention deteriorates

Engineering Contradiction:

Improveline-of-sight prediction accuracyVSAvoidinterpretability of attention guidance factors

Core Design Contradiction:

Measurement precisionVSLoss of information

Solution Approach 1:

The patent segments the complex deep learning model into distinct functional components: low-order feature extraction module, high-order feature extraction module, and saliency calculation module. Each module processes specific types of features (color, luminance, bearing for low-order; processing fluency, position bias for high-order), making the attention guidance factors interpretable while maintaining prediction accuracy through systematic feature integration.

Inventive Principle:

Principle #1Segmentation

2Device complexity

If only low-order image feature amounts are used, then computational simplicity is maintained, but line-of-sight guidance predictability deteriorates

Engineering Contradiction:

Improvecomputational simplicityVSAvoidline-of-sight guidance predictability

Core Design Contradiction:

Device complexityVSMeasurement precision

Solution Approach 1:

The patent merges low-order image feature amounts (color, luminance, bearing) with high-order image feature amounts (processing fluency, position bias) in a unified saliency calculation framework. This combination integrates both simple computational features and complex cognitive features, achieving accurate line-of-sight guidance predictability while maintaining reasonable computational complexity through efficient feature fusion.

Inventive Principle:

Principle #5Merging (Combining)

3Measurement precision

If deep learning models are used for saliency evaluation, then evaluation accuracy is improved, but ease of understanding for designers deteriorates

Engineering Contradiction:

Improvesaliency evaluation accuracyVSAvoidunderstandability for designers

Core Design Contradiction:

Measurement precisionVSEase of operation

Solution Approach 1:

The patent implements feedback mechanisms that provide designers with interpretable information about which specific factors (low-order features like color and luminance, high-order features like processing fluency and position bias) are guiding attention in the saliency evaluation results. This feedback loop allows designers to understand and adjust their designs based on clear, factor-specific insights while maintaining high evaluation accuracy.

Inventive Principle:

Principle #23Feedback

Data Source

PatentEP4287139B1Saliency analysis system, saliency analysis method and program

Publication Date: 2026.04.29 KONICA MINOLTA INC

Figure 1
Figure 2A~2D
Figure 3

AI summary

A saliency analysis system, includes: an input receiver that receives an evaluation target image; a feature amount extractor that extracts low-order image feature amounts and high-order image feature amounts, from the evaluation target image, and a calculator that calculates saliencies in the image, based on the low-order image feature amounts and the high-order image feature amounts.