Image Recognition Using Gaussian Probability Mapping for Pair Separation

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing image recognition methods, particularly in face and fingerprint recognition, struggle with non-linear variations and complex decision boundaries, leading to false positives and inadequate separation of matching and non-matching image pairs.

Innovation Solution

A neural network is configured to map input image data to predefined Gaussian probability distributions, using a Siamese neural network for feature extraction and a convolutional neural network for metric learning, minimizing Kullback-Leibler divergence to separate matching and non-matching pairs effectively.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If conventional distance measures (l2) and loss functions (softmax cross-entropy, contrastive loss, triplet loss) are used to compute distance between embeddings, then the system can perform basic face recognition, but the decision boundaries become complex and arbitrary leading to false positives and inadequate separation of matching and non-matching pairs

Engineering Contradiction:
Improveaccuracy of face recognitionVSAvoidcomplexity of decision boundaries
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent transforms the output of the neural network from raw embedding distances to probability values through a sigmoid function, changing the parameter space from Euclidean distance to probability space. This transformation simplifies the decision boundary to a fixed threshold (0.5) while maintaining high accuracy in distinguishing matching from non-matching face pairs.

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent introduces a sigmoid function as an intermediary between the embedding distance computation and the final decision. This intermediary transforms the complex distance relationships into a simplified probability output, making the decision boundary explicit and fixed rather than complex and arbitrary.

Inventive Principle:
Principle #24Intermediary (Mediator)

2Reliability

If angular distance metrics are used to enforce large margins between negative examples, then the number of false positives is reduced, but the training process becomes more complex and requires learning on a hypersphere manifold

Engineering Contradiction:
Improvereduction of false positivesVSAvoidcomplexity of training on hypersphere manifold
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

Instead of directly learning angular embeddings on a hypersphere manifold, the patent copies the successful distance separation achieved by angular metrics and reconstructs it in probability space using a simple sigmoid function. This avoids the complexity of hypersphere learning while preserving the false positive reduction benefit.

Inventive Principle:
Principle #26Copying

Solution Approach 2:

The patent replaces the complex mechanical system of angular distance computation and hypersphere manifold learning with a simpler probabilistic model using the sigmoid function. This substitution maintains the ability to separate positive and negative examples while eliminating the need for complex angular computations during training and inference.

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

3Productivity

If the system uses predefined analytical metrics to compute distance between embeddings, then the recognition process is computationally efficient, but the system struggles with non-linear variations and non-ideal conditions in face images

Engineering Contradiction:
Improvecomputational efficiencyVSAvoidrobustness to non-linear variations
Core Design Contradiction:
ProductivityVSReliability

Solution Approach 1:

The patent performs preliminary non-linear feature extraction using a trained neural network before the distance computation step. This preliminary action captures complex non-linear variations in face images, allowing the subsequent simple sigmoid-based decision to remain computationally efficient while being robust to non-ideal conditions.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent segments the face recognition process into two distinct stages: (1) complex non-linear feature extraction through neural network embeddings, and (2) simple probabilistic decision-making through sigmoid function. This segmentation allows each stage to specialize - the neural network handles non-linear variations while the sigmoid function maintains computational efficiency.

Inventive Principle:
Principle #1Segmentation

Data Source

PatentEP3944138B1Method and apparatus for image recognition
Publication Date: 2026.02.11 SONY SEMICON SOLUTIONS CORP
  • EP3944138B1 patent drawingFigure 1
  • EP3944138B1 patent drawingFigure 2
  • EP3944138B1 patent drawingFigure 3

AI summary

The present disclosure relates to an apparatus for image recognition. The apparatus comprises a machine learning network configured to map first and second input image data to either a first or a second predefined target probability distribution, depending on whether the first and second input image data correspond to matching or non-matching images, wherein an output of the machine learning network matching the first target probability distribution is indicative of matching images and an output of the machine learning network matching the second target probability distribution is indicative of non-matching images. The present disclosure also relates to a method for training the apparatus for image recognition.