Convolutional Neural Network Shared Feature Extraction for Imbalanced Category Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image processing techniques face challenges in accurately classifying image data into multiple categories using convolution neural networks, particularly when there is an imbalance in learning data between categories, leading to erroneous distinctions.
Innovation Solution
The proposed solution involves an image processing apparatus with a convolution neural network that performs convolution and total combination processes using learned coefficients, and a normalization unit to determine the category of input image data, while allowing for flexible category addition or deletion without increasing computational or memory resources.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If multiple neural networks are used for category classification, then classification accuracy for multiple categories improves, but device complexity and computational resources increase
Solution Approach 1:
The patent merges multiple category-specific neural networks into a single unified neural network that processes all categories simultaneously. The network uses a shared feature extraction layer followed by category-specific classification layers, combining the functionality of multiple networks while reducing overall complexity and resource requirements.
Solution Approach 2:
The neural network is designed with universal feature extraction capabilities that serve all category classifications. The shared convolutional layers and pooling layers perform multi-functional feature detection that benefits all category types, eliminating the need for separate specialized networks for each category.
2Reliability
If more learning data is collected for all categories, then classification reliability improves, but data collection time and storage requirements increase
Solution Approach 1:
The patent changes the parameter of data utilization by implementing data augmentation techniques that generate synthetic training samples through transformations such as rotation, scaling, and color adjustments. This approach increases the effective size and diversity of the training dataset without requiring additional physical data collection, thereby improving classification reliability while avoiding time losses.
3Measurement precision
If data augmentation is performed, then classification accuracy with limited data improves, but processing time increases
Solution Approach 1:
The patent applies partial data augmentation by selectively augmenting only certain types of training data or applying augmentation to a subset of the training process. This approach provides sufficient improved classification accuracy for limited datasets while avoiding the excessive processing time that would result from augmenting all training data comprehensively.
4Measurement precision
If category-specific processing is implemented, then classification precision for each category improves, but device complexity increases
Solution Approach 1:
The patent segments the neural network into distinct functional modules: shared feature extraction layers, category-specific classification layers, and output layers. This segmentation allows each category to have specialized processing while maintaining overall system efficiency and reducing complexity compared to completely separate networks for each category.
Data Source
AI summary
An image processing apparatus calculating to determine one category, to which an input image data belongs, among categories, which are provided to classify image data, includes a convolution processing unit which performs a convolution process using a first coefficient used for calculating to determine the one category and previously learnt; total combination processing units which are prepared for each category performs a total combination process for a processing result obtained by the convolution processing unit using a second coefficient used for calculating to determine the one category and previously learnt, and calculates to determine the one category; and a normalization unit which performs normalization of a processing result obtained by the total combination processing unit for each category, wherein the convolution processing unit performs learning of the first coefficient in common with the categories, and the total combination processing unit performs learning of the second coefficient for each category.