Image Analysis Device Context Shape Score Correction

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

In autonomous driving, existing image analysis systems using neural networks for context and shape analysis perform suboptimally due to the lack of integration and sharing of convolutional features between different networks, leading to inefficient analysis performance.

Innovation Solution

An image analysis device that combines context and shape analysis by using a convolutional neural network to calculate feature maps, which are then used to generate context and shape score information, with a correction unit that enhances shape score accuracy by multiplying shape scores with context scores, thereby improving the integration and accuracy of analysis.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If separate neural networks are used for context analysis and shape analysis, then each network can be optimized for its specific task, but the analysis performance does not improve due to lack of integration and feature sharing

Engineering Contradiction:
Improveanalysis performanceVSAvoidnetwork integration
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent combines context analysis network and shape analysis network into a unified architecture where both networks share common convolutional feature extraction components. This merging allows feature sharing and integration while maintaining specialized processing paths, resolving the contradiction between improved analysis performance and system complexity

Inventive Principle:
Principle #5Merging (Combining)

Solution Approach 2:

The common convolutional layers serve multiple functions by providing shared feature representations for both context analysis and shape analysis tasks. This multi-functionality enables both specialized networks to benefit from the same extracted features, improving overall analysis performance without proportionally increasing complexity

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

2Productivity

If convolutional features are not shared between networks, then each network maintains independence, but feature utilization efficiency decreases leading to suboptimal analysis performance

Engineering Contradiction:
Improvefeature utilization efficiencyVSAvoidanalysis performance
Core Design Contradiction:
ProductivityVSMeasurement precision

Solution Approach 1:

By merging the feature extraction pathways and implementing shared convolutional layers between context and shape analysis networks, the system achieves efficient feature utilization. The same computational features are reused across both analysis tasks, improving productivity without sacrificing analysis precision

Inventive Principle:
Principle #5Merging (Combining)

Data Source

PatentEP3410344B1Image analysis device, image analysis method, and computer-readable medium
Publication Date: 2023.03.01 KK TOSHIBA
  • EP3410344B1 patent drawingFigure 1
  • EP3410344B1 patent drawingFigure 2~3
  • EP3410344B1 patent drawingFigure 4~5

AI summary

According to an arrangement, an image analysis device (10) includes a feature amount calculation unit (14), a context calculation unit (16), a shape calculation unit (20), a correction unit (24), and an output unit (26). The feature amount calculation unit (14) is configured to calculate a feature map of a target image. The context calculation unit (16) is configured to calculate context score information representing context of each pixel on the basis of the feature map. The shape calculation unit (20) is configured to calculate shape score information representing a shape of an object in at least one region in the target image on the basis of the feature map. The correction unit (24) is configured to correct the shape score information in the at least one region using the context score information. The output unit (26) is configured to output the corrected shape score information.