Salient Object Detection With Difference Convolution for Real-Time Video

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

High computational costs of machine learning models for salient object detection in images and videos make them unsuitable for deployment on resource-constrained devices and real-time processing, despite their high accuracy.

Innovation Solution

A lightweight machine learning model utilizing spatial temporal difference convolution blocks and difference convolution reparameterization techniques to efficiently extract spatial and temporal features without increasing the number of parameters, enhancing model capacity for real-time inference.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If machine learning models with millions of parameters are used for salient object detection, then detection accuracy is improved, but computational cost increases making them unsuitable for resource-constrained devices and real-time processing

Engineering Contradiction:
Improvedetection accuracyVSAvoidcomputational cost
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The model segments feature extraction into spatial and temporal components using separate convolution blocks. Spatial convolution blocks process spatial features while temporal convolution blocks process temporal features, dividing the complex detection task into manageable parts that reduce overall computational burden while maintaining accuracy

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent extracts only the essential spatial and temporal features needed for salient object detection using difference convolution operations. By taking out and processing only the relevant features rather than all possible features, the model achieves accurate detection with reduced computational cost compared to full-model approaches

Inventive Principle:
Principle #2Taking out (Extraction)

2Reliability

If model capacity is increased to improve detection performance, then accuracy is improved, but the number of parameters increases requiring more computational resources

Engineering Contradiction:
Improvedetection performanceVSAvoidnumber of parameters
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent changes the parameter representation by using difference convolution operations that compute feature differences rather than full convolution. This parameter transformation allows the model to achieve higher detection performance through more efficient feature representation without proportionally increasing the number of parameters

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The spatial and temporal difference convolution blocks serve multiple functions: they extract spatial features, extract temporal features, and implicitly capture motion information. This multi-functionality allows the model to achieve high detection performance without needing separate dedicated components for each function, thereby avoiding parameter inflation

Inventive Principle:
Principle #6Universality (Multi-functionality)

Data Source

PatentEP4517682B1Real time salient object detection in images and videos
Publication Date: 2026.04.15 INTEL CORP
  • EP4517682B1 patent drawingFigure 1
  • EP4517682B1 patent drawingFigure 2
  • EP4517682B1 patent drawingFigure 3A~3D

AI summary

Machine learning models can generate outputs such as salient object detection. However, top-performing large-scale models have high computational cost, making the models hard to deploy on resource-constrained devices and to make predictions in real time. A difference convolution reparameterization technique can be used to fuse the sets of parameters learned during training time into a single set of parameters to be used at inference time or at deployment. The technique can increase performance without adding a lot of parameters. Furthermore, an efficient and effective spatial temporal difference convolution block can be used to leverage temporal information to make predictions. The spatial temporal difference convolution block can also apply the difference convolution reparameterization technique to reduce the need to increase parameters to achieve desired performance.