Learning Data Generating System Using Feature Map Subtraction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional deep learning methods for image recognition, such as Manifold Mixup, face challenges in preserving texture information when combining feature maps of two images, leading to inadequate learning of subtle differences in textures, particularly in medical images like ultrasonic endoscope images, resulting in insufficient recognition accuracy.
Innovation Solution
A learning data generating system that replaces parts of the feature maps from one image with corresponding parts of another image, preserving texture information without weighting addition, allowing for improved recognition accuracy even with limited data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If feature maps are combined by weighted addition (Manifold Mixup), then learning data is padded out, but texture information is lost
Solution Approach 1:
Instead of combining feature maps through weighted addition (forward approach), the patent inverts the approach by performing weighted subtraction to generate pseudo-feature maps. This inversion preserves texture information while still achieving data augmentation, as the subtraction operation maintains the structural characteristics of the original feature maps better than addition does.
Solution Approach 2:
The patent changes the operational parameter from addition to subtraction when combining feature maps. By using weighted subtraction instead of weighted addition, the method maintains texture information integrity while still generating diverse learning data, effectively resolving the contradiction between data quantity and information preservation.
2Quantity of substance
If feature maps are combined by weighted addition, then learning data is padded out, but recognition accuracy for subtle texture differences is insufficient
Solution Approach 1:
The patent inverts the conventional combination operation from addition to subtraction. This inversion enables the generation of pseudo-feature maps that preserve subtle texture differences, thereby improving recognition accuracy while still achieving the goal of padding learning data.
Solution Approach 2:
The patent substitutes the mechanical operation of weighted addition with weighted subtraction. This operational substitution fundamentally changes how feature maps are combined, allowing preservation of fine texture details that are crucial for accurate recognition of subtle differences in medical images.
3Productivity
If original learning data is used as basis for padding, then data efficiency is improved, but texture information preservation is insufficient
Solution Approach 1:
The patent changes the operational parameter from addition-based combination to subtraction-based combination. This parameter change allows the method to efficiently pad learning data using original data as basis while simultaneously preserving texture information that would otherwise be lost in conventional addition-based methods.
Data Source
AI summary
A learning data generating system includes a processor. The processor inputs a first image to a first neural network to generate a first feature map by the first neural network and inputs a second image to the first neural network to generate a second feature map by the first neural network. The processor generates a combined feature map by replacing a part of the first feature map with a part of the second feature map. The processor inputs the combined feature map to a second neural network to generate output information by the second neural network. The processor calculates an output error based on output information, first correct information, and second correct information