Virtual Timber Board Dataset Generation for Neural Network Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The lack of a comprehensive training dataset for artificial neural networks to accurately detect characteristics of timber boards, such as pith location and defects, hinders their implementation in industrial settings, as each type of timber and growing conditions require specific training, and existing methods are inaccurate or impractical for industrial use.
Innovation Solution
A method to generate a training dataset for artificial neural networks using virtual models of timber boards, simulating real log structures and defects, which are then virtually sawn to create input-output data pairs for training, utilizing stochastic models to replicate the characteristics of different timber species and growing conditions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If actual timber boards with known characteristics are used to create training datasets, then the training data accuracy is improved, but the cost and time required to collect and annotate real timber data increases significantly
Solution Approach 1:
The patent creates virtual copies of timber boards through 3D modeling and rendering instead of using physical timber boards. Virtual timber boards are generated with realistic textures, knots, defects, and growth ring patterns, then rendered from multiple camera angles to create training images. This copying approach maintains measurement precision while eliminating the time-consuming process of collecting, preparing, and annotating physical timber samples.
2Productivity
If virtual models of timber boards are used to generate training datasets, then the data generation speed is improved, but the realism and accuracy of the training data may deteriorate
Solution Approach 1:
The patent applies local quality by incorporating specific realistic features into targeted regions of the virtual timber boards. Realistic wood grain patterns, knot structures, defect characteristics, and growth ring variations are added to specific local areas of the virtual models. The rendering process also applies localized lighting, shadows, and surface textures to enhance realism in critical regions while maintaining overall generation speed.
3Measurement precision
If comprehensive training datasets covering all timber types and conditions are created, then the neural network's detection accuracy is improved, but the complexity and size of the training dataset increases
Solution Approach 1:
The patent creates a universal virtual timber board generation system that can produce diverse training data across multiple timber types, defects, and conditions using a single platform. The virtual model system incorporates variable parameters for different wood species, knot types, defect characteristics, and environmental conditions, allowing one system to generate comprehensive training datasets for multiple detection scenarios without requiring separate physical timber collections for each case.
Data Source
Figure 1
Figure 2
Figure 3(a)~3(b)
AI summary
A computer-implemented method for generating a training dataset for training an artificial neural network configured to use images of lateral faces of a timber board to provide information about structure and/or defects, the method including: a log generation step during which a virtual model of a log is generated; a sawing step of the virtual model to obtain one or more virtual timber boards; a pattern step during which a surface pattern is determined as the intersection between the virtual lateral face and the internal structure and/or defects; a rendering step during which a rendered surface image of the lateral face of the virtual timber board is created; and an input data generation step during which the rendered surface images are used to create one or more item of input data; an output data generation step during which an item of output data is generated; and a population step during which a record is added to the training dataset comprising the item of input data, in combination with the item of output data.