Multivariate Data Imagification via Multi Area Colorization

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing methods for imagification of multivariate data sequences, such as those proposed in EP3239897A1, suffer from information loss, particularly in retaining vector direction and symmetry, leading to potential biases and increased model complexity due to fixed color encoding and the use of the L2 norm, which can degrade classification accuracy.

Innovation Solution

A Multi Area Colorization Technique is employed, where multivariate data sets are transformed into images by creating superimposed line graphs, colorizing areas under and above the graphs based on variable values relative to their means, and summing overlapping pixel colors, allowing for detailed representation of data evolution without information loss.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Ease of manufacture

If the L2 norm is used to encode plot height in imagification, then the transformation is mathematically simple, but information on vector direction is lost and square bias is introduced between training and test datasets

Engineering Contradiction:
Improvesimplicity of transformationVSAvoidloss of vector direction information
Core Design Contradiction:
Ease of manufactureVSLoss of information

Solution Approach 1:

The patent segments the multivariate data representation into multiple independent line graphs, each representing a different variable. This segmentation allows each variable to be visualized with its own color-coded line, preserving the directional information of each variable separately rather than losing it in a aggregated L2 norm calculation.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent uses color as an additional dimension to encode variable identity and direction information. By assigning unique colors to different variables and their positive/negative directions, the system preserves directional information that would otherwise be lost in traditional 2D plotting, effectively adding a dimensional layer to the visualization.

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

2Ease of manufacture

If fixed color encoding is used in imagification, then the implementation is straightforward, but model complexity increases and classification accuracy degrades

Engineering Contradiction:
Improveease of implementationVSAvoidclassification accuracy
Core Design Contradiction:
Ease of manufactureVSManufacturing precision

Solution Approach 1:

The patent applies local quality by assigning different color properties to different regions of the plot based on the specific variable and its direction. Each variable gets a unique color scheme where hue represents variable identity and saturation or lightness represents the magnitude and direction (positive/negative), allowing the visualization to adapt locally to the data characteristics rather than using a fixed global color scheme.

Inventive Principle:
Principle #3Local quality

Solution Approach 2:

The patent systematically uses color changes to encode multiple dimensions of information. By varying hue, saturation, and lightness based on variable identity, direction, and magnitude, the system conveys rich information through color alone, reducing the need for complex additional visual elements and improving classification accuracy while maintaining implementation simplicity.

Inventive Principle:
Principle #32Color changes

3Shape

If symmetric plot shape is used about the x-axis, then the visualization is aesthetically balanced, but information is effectively duplicated and model training time increases

Engineering Contradiction:
Improvesymmetry of plotVSAvoidmodel training time
Core Design Contradiction:
ShapeVSLoss of time

Solution Approach 1:

The patent introduces asymmetry by plotting variables above and below the x-axis differently, with distinct color coding for positive and negative directions. This asymmetric approach eliminates redundant information duplication while maintaining visual clarity, as each side of the x-axis now carries unique directional information rather than mirroring the same data.

Inventive Principle:
Principle #4Asymmetry

4Loss of information

If areas under line graphs are colorized based on variable values, then detailed data evolution is represented, but overlapping areas require color summation which complicates the process

Engineering Contradiction:
Improveretention of data detailsVSAvoidcomplexity of colorization process
Core Design Contradiction:
Loss of informationVSDevice complexity

Solution Approach 1:

The patent uses parameter changes in the color model (hue, saturation, lightness) to encode different variables and their values. By systematically varying these color parameters based on variable identity and magnitude, the system can represent detailed data evolution through simple color assignments rather than complex color mixing, even in overlapping areas.

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentEP3588441B1Imagification of multivariate data sequences
Publication Date: 2023.08.09 FUJITSU LTD
  • EP3588441B1 patent drawingFigure 1~3(b)
  • EP3588441B1 patent drawingFigure 4
  • EP3588441B1 patent drawingFigure 5

AI summary

A digital image comprising superimposed line graphs corresponding to different variables in a multivariate data set is created, the graphs representing values of the variables in successive items of the data set. The digital image is colorized according to a process in which, for each positive value variable, a color is assigned to an area under the line graph in dependence upon a color pre-allocated to that variable and a mean value of the data item. For each negative value variable in the data item, a color is assigned to an area above the line graph in dependence upon a color pre-allocated to that variable and a mean value of the data item. The colors where the areas for at least two of the line graphs overlap are summed. This method allows more information from the data set to be captured in the image, which may be used in a neural network trained for image recognition.