Object Recognition Training Data Verification for Label Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional methods for evaluating labeled training data in machine learning are time-consuming and labor-intensive, especially with large datasets, affecting the accuracy of object recognition models.
Innovation Solution
A method involving data clustering and optimization techniques to group and refine training areas and properties, using algorithms like DBSCAN, to ensure accurate alignment of training data with objects, followed by model building and recognition using machine learning algorithms like CenterNet, YOLO, and R-CNN.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual evaluation is used to verify labeled training data, then accuracy of data labeling can be ensured, but time consumption and labor intensity increase significantly
Solution Approach 1:
The patent introduces an automated verification system that acts as an intermediary between data labelers and model trainers. This system uses pre-trained models to automatically verify labeled data, replacing manual evaluation while maintaining accuracy. The verification system processes large volumes of data quickly, resolving the contradiction between evaluation accuracy and time consumption.
Solution Approach 2:
The patent replaces the mechanical manual evaluation process with an automated computational system. Instead of human reviewers manually checking each label, the system uses algorithmic verification methods including consistency checks, cross-validation with pre-trained models, and automated anomaly detection. This substitution dramatically reduces time consumption while maintaining or improving verification accuracy.
2Reliability
If manual evaluation is used to verify labeled training data, then data quality can be controlled, but labor intensity increases significantly
Solution Approach 1:
The patent implements a self-verification mechanism where the system automatically checks and validates labeled data without requiring manual intervention. The verification system performs consistency checks, detects anomalies, and flags problematic labels automatically. This self-service approach maintains data quality control while eliminating the need for extensive manual labor, thereby improving ease of operation.
Solution Approach 2:
The patent incorporates feedback loops where the automated verification system continuously monitors data quality metrics and provides real-time feedback to the labeling process. When issues are detected, the system automatically adjusts verification parameters or flags specific labels for review. This feedback mechanism ensures consistent data quality control while reducing overall labor intensity through automated detection and correction.
3Productivity
If clustering algorithms are used to group training areas, then data processing efficiency improves, but computational complexity increases
Solution Approach 1:
The patent applies segmentation by dividing the large training dataset into smaller clusters using clustering algorithms. Instead of processing all training data uniformly, the system groups similar training areas together, allowing for more efficient processing of each cluster. This segmentation improves productivity by enabling parallel processing and reducing the computational burden on individual processing units, while the modular cluster structure helps manage computational complexity.
Data Source
AI summary
A method for evaluating data to be used to train an object recognition model is to be implemented by a computing device. The computing device stores a plurality of training datasets respectively related to a plurality of images, and each training dataset includes a plurality of entries of training data. The method includes steps of: obtaining, for each image, at least one target area and at least one target property that are related to the image based on the entries of training data; creating, for each image, a training material that includes the image, and the at least one target area and the at least one target property both related to the image; and obtaining at least three object recognition models based on the training materials that are created respectively for the images using at least one machine learning algorithm.