Pre-Trained Neural Network Adaptation for Image Analysis

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing deep neural networks require extensive annotated training data for updates and retraining when new data or changing conditions are encountered, which is time-consuming and costly, especially in applications like surveillance and autonomous navigation.

Innovation Solution

A method using a pre-trained deep neural network adapted for a specific class within a hierarchical system, allowing for online updates with less accurate annotations and focusing on features corresponding to the specific class through a forward and optional backward pass, enabling efficient analysis of new image data without extensive retraining.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If a deep neural network is retrained or updated with new data, then the model accuracy and adaptability to new conditions improve, but the time consumption and cost increase significantly due to required annotations

Engineering Contradiction:
Improvemodel adaptability to new data and conditionsVSAvoidtime consumption for annotation and retraining
Core Design Contradiction:
Adaptability or versatilityVSLoss of time

Solution Approach 1:

The system performs preliminary action by using a pre-trained neural network as a starting point. Instead of training from scratch or requiring full retraining, the pre-trained model's weights and architecture are leveraged, allowing rapid adaptation to new data with minimal annotation effort. This preliminary preparation eliminates the need for time-consuming retraining while maintaining adaptability.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The invention applies copying by creating a copy of the pre-trained neural network model and adapting it to new conditions. Rather than reconstructing the model from scratch, a copy of the existing model is modified with new data, preserving the learned features while adapting to new scenarios. This copying approach significantly reduces annotation and retraining time.

Inventive Principle:
Principle #26Copying

2Reliability

If a deep neural network is trained for a specific task, then the model performance for that task improves, but the ability to perform different related tasks without retraining is limited

Engineering Contradiction:
Improvemodel performance for specific taskVSAvoidability to perform different related tasks
Core Design Contradiction:
ReliabilityVSAdaptability or versatility

Solution Approach 1:

The system achieves universality by designing a pre-trained neural network that can serve multiple related tasks. The pre-trained model captures general features and patterns that can be applied across different tasks, allowing the same model to perform various related functions by simply changing the task parameters or data, without requiring retraining. This multi-functionality maintains high performance across different tasks.

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

Solution Approach 2:

The invention uses parameter changes to adapt the pre-trained model to different tasks. By modifying parameters such as loss functions, learning rates, or task-specific layers while keeping the core architecture and pre-trained weights intact, the model can efficiently adapt to new related tasks. This parameter adjustment approach preserves task-specific performance while enabling versatility.

Inventive Principle:
Principle #35Parameter changes

3Measurement precision

If new data from target site becomes available, then the completeness and accuracy of analysis improves, but the requirement for annotation and model update increases the complexity and cost

Engineering Contradiction:
Improveanalysis accuracy with new dataVSAvoidcomplexity of annotation and model update process
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The system applies self-service by enabling the pre-trained neural network to automatically adapt to new data without requiring extensive manual annotation. The model uses its own pre-learned features and architecture to process new data, performing self-adjustment through techniques like fine-tuning or continued learning. This self-service capability maintains analysis accuracy while significantly reducing the complexity and cost of updates.

Inventive Principle:
Principle #25Self-service

Solution Approach 2:

The invention introduces an intermediary approach by using the pre-trained model as a mediator between new data and the analysis system. Instead of directly processing raw new data requiring full retraining, the pre-trained model acts as an intermediary that transforms and adapts the new data, simplifying the update process while maintaining measurement precision.

Inventive Principle:
Principle #24Intermediary (Mediator)

Data Source

PatentEP3596655B1Method and apparatus for analysing an image
Publication Date: 2023.08.09 SIEMENS AG
  • EP3596655B1 patent drawingFigure 1~2
  • EP3596655B1 patent drawingFigure 3~5
  • EP3596655B1 patent drawingFigure 6~7

AI summary

The invention relates to a method (1, 22) and an apparatus for analysing an image (16, 23, 43). The method (1, 22) comprises providing an artificial deep neural net (2, 24) pre-trained to classify images (16, 23, 43) into a hierarchical system of multiple hierarchical classes. The pre-trained neural net (2, 24) is then adapted for one specific class, wherein the specific class is lower in the hierarchical system than an actual class of the image (16, 23, 43). The image (16, 23, 43) is then processed by means of a forward pass (26) through the adapted neural net (5) to generate a processing result (7, 20). An image processing algorithm (28) is then used to analyse the processing result (7, 20) focused on features (21) corresponding to the specific class.