Capsule Neural Networks for Spatially Accurate Multi-Task Vision

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Deep artificial neural networks struggle with accurately identifying spatial positions of features in images, particularly when images are tilted or tampered, and fail to effectively capture interdependent properties of input data, leading to inaccurate segmentation and classification tasks.

Innovation Solution

A method for training artificial neural networks that includes a capsule layer configured to output pose arrays, where the network is trained with labeled samples to optimize a loss function that penalizes differences between ground truth labels and generated outputs, allowing the network to learn interpretable features and solve multiple tasks simultaneously.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If deep artificial neural networks are used to identify features in images, then feature detection capability is improved, but spatial position accuracy and robustness to image transformations deteriorate

Engineering Contradiction:
Improvefeature detection accuracyVSAvoidspatial position accuracy
Core Design Contradiction:
Measurement precisionVSReliability

Solution Approach 1:

The patent transforms the output of neurons from scalar values to pose vectors that encode spatial position, orientation, and scale information. This dimensional expansion allows the network to simultaneously detect features and accurately represent their spatial properties, resolving the contradiction between feature detection and spatial accuracy.

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

Solution Approach 2:

The patent changes the parameter representation from simple activation values to structured pose parameters (position x, y; orientation θ; scale s). This parameter transformation enables the network to maintain accuracy under image transformations by explicitly modeling spatial relationships rather than relying on fixed-position detectors.

Inventive Principle:
Principle #35Parameter changes

2Measurement precision

If more training data with data augmentation is used to recognize features in tilted images, then recognition accuracy is improved, but training complexity and computational resources increase

Engineering Contradiction:
Improvefeature recognition accuracyVSAvoidtraining complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent performs preliminary action by pre-computing pose transformations during training. Instead of relying on large amounts of augmented training data, the network learns to predict pose parameters that inherently account for transformations, reducing the need for extensive data augmentation and simplifying the training process.

Inventive Principle:
Principle #10Preliminary action

3Loss of information

If capsule networks with routing procedures are used to preserve part-whole relationships, then information preservation is improved, but network complexity increases

Engineering Contradiction:
Improvepart-whole relationship informationVSAvoidnetwork structure complexity
Core Design Contradiction:
Loss of informationVSDevice complexity

Solution Approach 1:

The patent applies segmentation by dividing the network into distinct capsule layers where each layer processes specific hierarchical levels of features. This structured segmentation preserves part-whole relationships by organizing neurons into capsules that represent objects and their parts at different scales, making the complex information flow more manageable and interpretable.

Inventive Principle:
Principle #1Segmentation

Data Source

PatentEP4066167B1Solving multiple tasks simultaneously using capsule neural networks
Publication Date: 2026.03.04 LARALAB UG
  • EP4066167B1 patent drawingFigure 1~2
  • EP4066167B1 patent drawingFigure 3
  • EP4066167B1 patent drawingFigure 4~5

AI summary

The invention provides a system and method for training artificial neural networks for solving multiple tasks simultaneously, wherein the artificial neural network comprises at least one capsule layer. The invention also provides a system and a method for solving multiple tasks simultaneously, wherein the artificial neural network comprises at least one capsule layer. The invention further provides additional connected aspects.