Weighted Spectral Saliency for Low-SWaP Object Detection

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing object detection systems struggle to accurately identify salient objects in wide-area images while minimizing false alarms and computational overhead, particularly on low size, weight, and power (SWaP) platforms.

Innovation Solution

A weighted spectral saliency system that uses a learned weight matrix to amplify foreground object saliency signatures and suppress background signatures, integrating knowledge from annotated training sets to reduce false alarms and computational load.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If learning-based object detection systems are used to accurately detect salient objects, then detection accuracy is improved, but computational requirements increase significantly

Engineering Contradiction:
Improvedetection accuracyVSAvoidcomputational requirements
Core Design Contradiction:
Measurement precisionVSPower

Solution Approach 1:

The patent introduces spectral saliency computation as an intermediary step between image input and object detection. This intermediary computes a saliency map that highlights potential object regions, which then guides the learning-based detector to focus only on these regions rather than processing the entire image, thereby reducing computational load while maintaining detection accuracy

Inventive Principle:
Principle #24Intermediary (Mediator)

Solution Approach 2:

The patent segments the image processing task into two distinct stages: first computing spectral saliency to identify potential object regions, then applying learning-based detection only to these segmented regions. This segmentation allows the system to leverage the strengths of both spectral methods (low computational cost for broad coverage) and learning-based methods (high accuracy for specific objects)

Inventive Principle:
Principle #1Segmentation

2Productivity

If conventional spectral saliency systems are used to detect objects in wide-area images, then computational efficiency is improved, but false alarms from background objects increase

Engineering Contradiction:
Improvecomputational efficiencyVSAvoidfalse alarm rate
Core Design Contradiction:
ProductivityVSReliability

Solution Approach 1:

The patent implements feedback by using the spectral saliency map to guide the learning-based object detector. The saliency map provides initial region proposals that are then refined and validated by the learned detector, creating a feedback loop where the detector's results can further inform subsequent detection iterations, thereby reducing false alarms from background objects

Inventive Principle:
Principle #23Feedback

3Adaptability or versatility

If learning-based object detection systems are deployed on low SWaP platforms, then adaptability to constrained environments is improved, but system complexity increases

Engineering Contradiction:
Improveadaptability to low SWaP platformsVSAvoidsystem complexity
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The patent segments the computational workload into two parts: spectral saliency computation that can be performed with limited resources, and learning-based detection that operates only on reduced regions of interest. This segmentation makes the overall system adaptable to low SWaP platforms by avoiding the need to deploy full learning-based detection across the entire image

Inventive Principle:
Principle #1Segmentation

Data Source

PatentEP3635631B1System for detecting salient objects in images
Publication Date: 2025.12.03 HRL LAB
  • EP3635631B1 patent drawingFigure 1
  • EP3635631B1 patent drawingFigure 2
  • EP3635631B1 patent drawingFigure 3

AI summary

Described is a system for detecting salient objects in images. During operation, the system maps an input image into a frequency domain having a spectral magnitude. The spectral magnitude is replaced with weights from a weight matrix W. The frequency domain is then transformed with the weights to a saliency map in the image domain, the saliency map having pixels with pixel values. A squaring operation is then performed on the saliency map by squaring the pixel values to generate a pixel-value altered saliency map. A final saliency map is generated by filtering the pixel-value altered saliency map. A number of devices may then be operated based on the saliency map.