Multi-Scale ML Image Model for Fine Shape Reconstruction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning models, such as Variational Autoencoders (VAE), struggle to accurately represent fine shapes of objects in generated image data due to information compression, leading to loss of small features.
Innovation Solution
A machine learning model with multiple extraction and restoration units, including convolution and transposed convolution layers, processes input image data to generate output image data that represents both fine and coarse shapes by extracting and adding displacement information, improving robustness against position displacement.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of information
If information compression is used to extract features (e.g., VAE), then abstract features can be extracted, but small features and fine shapes are lost
Solution Approach 1:
The model segments the feature extraction process into multiple parallel extraction units, each targeting different scales of input ranges. The first extraction unit processes fine details (small input range), while the second extraction unit processes broader structures (larger input range). This segmentation allows the model to preserve both fine shapes and abstract features without information loss.
Solution Approach 2:
The patent introduces a new dimension of processing by adding displacement information obtaining units that operate in parallel with the extraction units. These units capture positional displacement information, adding a spatial dimension to the feature representation. This enables the model to recover fine shape information that would otherwise be lost in compression.
2Measurement precision
If multiple extraction units with different input ranges are used, then fine and coarse shapes can be captured, but device complexity increases
Solution Approach 1:
The model merges the outputs of multiple extraction units and displacement information obtaining units through addition operations. By combining these parallel processing paths, the model achieves comprehensive shape feature extraction (both fine and coarse) while maintaining a unified architectural framework that manages complexity through systematic integration.
3Reliability
If displacement information is added to improve robustness, then position displacement resistance improves, but calculation complexity increases
Solution Approach 1:
The displacement information obtaining units act as intermediaries between the extraction units and the final output. These units specifically capture positional displacement information and feed it into the addition unit, which integrates it with the extracted features. This intermediary approach improves robustness against position displacement by dedicating specific computational components to handle displacement information separately.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
The present invention generates output image data representing an output object having a fine shape. The machine learning model includes an input layer and an output layer. The input layer includes a first extraction unit, a second extraction unit, one or more displacement information obtaining units, and an addition unit. The first extraction unit extracts a shape feature of a portion included in a first input range of an input object. The second extraction unit is provided in parallel with the first extraction unit, and extracts a shape feature of a portion included in a second input range of the input object, the second input range being larger than the first input range. The one or more displacement information obtaining units obtain displacement information relating to position displacement within an input image of the shape feature extracted by the first extraction unit. The addition unit adds a plurality of pieces of data, including data outputted from the first extraction unit, data outputted from the second extraction unit, and data outputted from the one or more displacement information obtaining units.