Image Labeling Apparatus Using Clustering for Semi-Automatic Data Annotation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep learning models require large amounts of labeled data for effective training, which is costly and time-consuming to obtain, especially in industries like textiles where data diversity and defect categories are high, making it difficult to implement fully automatic labeling systems.
Innovation Solution
An image labeling apparatus and method using clustering technology that generates image features, classifies them into groups, selects representative features, and assigns labels semi-automatically, reducing the need for extensive manual labeling by experts.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual labeling by experts is used to ensure accurate classification of diverse image categories, then labeling accuracy is improved, but labeling cost and time consumption increase significantly
Solution Approach 1:
The system performs automatic clustering and representative feature selection to self-label the majority of images without expert intervention. The algorithm independently identifies defect categories, selects representative features, and assigns labels to image groups, enabling the system to serve its own labeling needs for routine cases.
Solution Approach 2:
The system pre-processes images by generating features, clustering them into groups, and selecting representative features before expert review. This preliminary automatic classification organizes the data structure and reduces the workload for experts, who only need to verify and confirm labels for representative images rather than examining every single image.
2Productivity
If fully automatic labeling system is implemented to reduce labeling cost, then labeling efficiency is improved, but adaptability to diverse product categories deteriorates
Solution Approach 1:
The system dynamically adapts to different product categories and defect types by learning from the data characteristics of each domain. The clustering algorithm automatically adjusts to the specific features and patterns present in different textile products, enabling the same system to handle diverse categories without manual reconfiguration while maintaining high labeling efficiency.
3Reliability
If expert review of every defective image is performed to ensure comprehensive defect categorization, then labeling completeness is improved, but processing speed decreases
Solution Approach 1:
The system segments the labeling task into two parts: automatic clustering and grouping of similar images, followed by expert verification only for representative features of each cluster. This segmentation allows the majority of images to be processed rapidly by the algorithm while experts focus their attention only on confirming the accuracy of cluster representations, thereby maintaining completeness without sacrificing speed.
Data Source
AI summary
An image labeling apparatus, method, and computer program product are provided. The image labeling apparatus generates several image features of several to-be-processed images and classifies them into several groups. For each group, the image labeling apparatus: (a) selects the image feature closest to the center as a representative feature and a compared feature, (b) selects the image feature farthest to the compared feature as a candidate feature, and (c) compares the similarity between the candidate feature and each representative feature with a threshold. When all the similarities are smaller than the threshold, adopts the candidate feature as another representative feature and the compared feature and repeats (b) and (c). The group(s) whose number of representative features is more than another threshold will be re-classified. The image labeling apparatus determines several to-be-labeled groups according to the classification result and labels the to-be-processed images within the same to-be-labeled group the same.