Convolutional Neural Network Filter Structure Determination

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

The determination of a convolutional neural network structure for image recognition is complex and typically requires expertise, making it difficult for non-experts to design an effective structure for image recognition tasks.

Innovation Solution

A method that involves acquiring initial filters with trained weights, splitting them using transformations from image processing fields, and merging filters through clustering to automatically determine the structure of a convolutional neural network, allowing for the adjustment of filter numbers and improving identification performance.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If manual design of convolutional neural network structure is performed by experts, then identification performance can be improved, but device complexity and difficulty of operation increase

Engineering Contradiction:
Improveidentification performanceVSAvoidstructure design complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The system performs automatic structure determination through self-organization of filters. Filters are automatically generated by applying transformations to initial filters and clustering them, eliminating the need for expert manual design while maintaining effective network structure for image recognition

Inventive Principle:
Principle #25Self-service

Solution Approach 2:

The invention changes the parameter of filter number from fixed expert-determined value to dynamically adjustable value. By controlling the number of clustering centers, users can easily adjust the network structure parameters without requiring expertise in neural network design

Inventive Principle:
Principle #35Parameter changes

2Ease of operation

If automatic structure determination is implemented, then ease of operation improves, but manufacturing precision may deteriorate

Engineering Contradiction:
Improveease of structure determinationVSAvoidstructure determination accuracy
Core Design Contradiction:
Ease of operationVSManufacturing precision

Solution Approach 1:

The system performs preliminary actions by generating multiple transformed filters from initial filters before clustering. This preliminary generation of diverse filter candidates ensures that the subsequent clustering process has high-quality inputs, maintaining determination accuracy while automating the process

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The system uses clustering evaluation to provide feedback on the quality of filter grouping. By evaluating clustering results and iteratively adjusting the number of clustering centers, the system ensures accurate structure determination while maintaining ease of operation

Inventive Principle:
Principle #23Feedback

3Reliability

If the number of filters is increased from N to M, then identification performance improves, but loss of time increases

Engineering Contradiction:
Improveidentification performanceVSAvoidtraining time
Core Design Contradiction:
ReliabilityVSLoss of time

Solution Approach 1:

The system performs preliminary training of initial filters before the splitting process. These pre-trained initial filters contain useful feature extraction capabilities that are then transformed and clustered, reducing the overall training time compared to training all filters from scratch

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The system creates transformed copies of initial filters through various image processing transformations rather than training entirely new filters. This copying approach with transformations preserves effective filter patterns while generating diversity, reducing training time while maintaining performance

Inventive Principle:
Principle #26Copying

Data Source

PatentEP3255606B1Determination method and program
Publication Date: 2020.10.07 PANASONIC INTELLECTUAL PROPERTY MANAGEMENT CO LTD
  • EP3255606B1 patent drawingFigure 1
  • EP3255606B1 patent drawingFigure 2A
  • EP3255606B1 patent drawingFigure 2B

AI summary

A determination method for determining the structure of a convolutional neural network includes the acquisition step of acquiring N filters having the weights trained using a training image group as the initial values, where N is a natural number greater than or equal to 1, (S10) and the splitting step of increasing the number of the filters from N to M, where M is a natural number greater than or equal to 2 and is greater than N, by adding a filter obtained by performing a transformation used in image processing fields on at least one of the N filters (S20).