Convolutional Neural Network Ensemble for Image Classification

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Conventional machine learning functions are difficult to adapt for pattern recognition in digital image processing, and determining the necessary architecture and parameters for convolutional neural networks (CNNs) for accurate image classification is challenging without a robust image processing system.

Innovation Solution

An iterative process is employed to build and fine-tune CNNs by selecting candidate architectures and parameters, evaluating performance, and generating an ensemble of the most accurate models to classify objects in digital images, with a focus on assessing property damage and enhancing image processing efficiency.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If conventional machine learning functions are used for pattern recognition in digital image processing, then the implementation is straightforward, but the classification accuracy and adaptability for complex image features are insufficient

Engineering Contradiction:
Improveclassification accuracyVSAvoidnetwork architecture complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent segments the image processing task into multiple convolutional layers, each performing specific feature detection at different levels of abstraction. The first layer detects basic features like edges and textures, while subsequent layers combine these to detect more complex patterns, enabling accurate classification without requiring a single monolithic complex model

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent introduces spatial dimensions through 2D convolutional kernels that operate across height and width of image features simultaneously. This dimensional approach allows the network to capture spatial relationships and patterns in images more effectively than conventional 1D or fully connected networks, improving accuracy while maintaining reasonable complexity

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

2Reliability

If the CNN architecture is customized for specific image processing tasks, then the classification accuracy improves, but the training time and computational resources increase significantly

Engineering Contradiction:
Improveimage classification accuracyVSAvoidtraining time
Core Design Contradiction:
ReliabilityVSLoss of time

Solution Approach 1:

The patent employs preliminary data normalization and feature preprocessing steps that standardize input images before they enter the CNN. This preliminary action reduces the computational burden during training by presenting consistently scaled and formatted data, thereby decreasing training time while maintaining the ability to achieve high accuracy with task-specific architecture customization

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent implements a modular CNN architecture where not all layers and components are activated or trained for every specific task. Instead, the network can selectively engage relevant layers based on the complexity of the classification task, reducing unnecessary computational overhead and training time while preserving accuracy for the specific application

Inventive Principle:
Principle #16Partial or excessive action

3Reliability

If multiple CNN models are trained and evaluated iteratively to find the optimal architecture, then the classification accuracy improves, but the development time and computational cost increase

Engineering Contradiction:
Improveensemble classification accuracyVSAvoidmodel development efficiency
Core Design Contradiction:
ReliabilityVSProductivity

Solution Approach 1:

The patent merges multiple CNN models into an ensemble architecture where individual models with different architectures and training parameters work together. By combining their predictions through voting or averaging mechanisms, the ensemble achieves superior classification accuracy that exceeds any single model, while the modular structure allows efficient development and deployment

Inventive Principle:
Principle #5Merging (Combining)

Data Source

PatentEP3065085B1Digital image processing using convolutional neural networks
Publication Date: 2020.04.01 ACCENTURE GLOBAL SERVICES LTD
  • EP3065085B1 patent drawingFigure 1
  • EP3065085B1 patent drawingFigure 2
  • EP3065085B1 patent drawingFigure 3

AI summary

According to an example, a digital image may be processed by an ensemble of convolutional neural networks (CNNs) to classify objects in the digital image. For each CNN, a candidate architecture and candidate parameters may be selected to build a plurality of CNNs. Once it is determined that a predetermined number of CNNs, each having different values for the selected candidate parameters, meet a validation threshold, an ensemble of CNNs may be generated from the predetermined number of CNNs. The predictions from the ensemble of CNNs may then be aggregated to accurately classify the objects in the digital image.