Neural Network Recognition Model Using Cluster-Specific Thresholds
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current machine recognition methods, particularly in artificial neural networks, face challenges in efficiently handling variations in object poses and ambient environments, leading to suboptimal performance in identifying and verifying input data such as images, fingerprints, or irises.
Innovation Solution
A processor-implemented recognition method that calculates feature data from input data using a recognition model, identifies clusters based on pose and environment, and trains neural networks with supervised and unsupervised learning to group and cluster data effectively, enabling accurate verification and identification across different conditions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional recognition methods are used to handle variations in object poses and ambient environments, then the system complexity remains low, but the recognition accuracy deteriorates
Solution Approach 1:
The patent segments the recognition process into multiple specialized neural networks: a first neural network for feature extraction, a second neural network for pose estimation, and a third neural network for environment recognition. This segmentation allows each network to specialize in specific aspects of the input data, improving overall recognition accuracy while managing system complexity through modular architecture.
Solution Approach 2:
The patent introduces cluster information as an additional dimension in the recognition process. By determining which cluster the input data belongs to (representing different pose or environment categories) and using this cluster information to select appropriate threshold values, the system enhances recognition accuracy without proportionally increasing complexity.
2Measurement precision
If cluster-based recognition with multiple neural networks is implemented, then recognition accuracy improves, but the training complexity increases
Solution Approach 1:
The training process is segmented into distinct phases for each neural network. The first neural network is trained on raw input data to extract features, the second network is trained on pose-related features, and the third network is trained on environment-related features. This segmented training approach reduces overall training complexity compared to training a single comprehensive network.
Solution Approach 2:
The system performs preliminary clustering of training data into groups representing different poses and environments before training the neural networks. This preliminary organization of data simplifies the training process by providing structured input patterns, allowing the networks to learn more efficiently from pre-organized clusters rather than raw unstructured data.
3Measurement precision
If cluster-specific threshold information is used for verification, then verification accuracy improves, but the processing time increases
Solution Approach 1:
Threshold values are pre-calculated and stored for each cluster during the training phase. During verification, the system only needs to determine the input's cluster membership and retrieve the corresponding pre-computed threshold, rather than calculating thresholds in real-time. This preliminary preparation significantly reduces verification processing time while maintaining high accuracy.
Solution Approach 2:
The system applies different threshold values locally to different clusters based on their specific characteristics. Each cluster has its own optimized threshold that is tailored to its particular pose or environment conditions, improving verification accuracy for each specific case without requiring a single conservative threshold that would be overly restrictive for all cases.
Data Source
AI summary
Provided is an apparatus and method for data recognition and training. The training apparatus generates a recognition model by grouping training data, performing clustering after excluding representative information from the grouped training data, and grouping the clustered data into sub-groups. The recognition apparatus identifies feature data and a cluster to which the feature data belongs from input data using a recognition model, and recognizes the input data based on the feature data and the cluster to which the feature data belongs.