Semantic Defect Embeddings for Dynamic Auto-Labeling
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing semi-supervised learning algorithms for defect detection are limited to handling known classes and fail to dynamically expand to accommodate new labels, and image feature embeddings are prone to noise due to redundant information, leading to inaccurate defect classification.
Innovation Solution
A feature extraction pipeline converts visual problems to textual problems by leveraging text feature embeddings, using a multi-modal machine learning model to generate structured text data and numerical embeddings, enabling precise defect identification and dynamic management of reference data structures.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If image feature embeddings are used for defect classification, then defect detection can be performed, but noise and redundant information lead to inaccurate classification
Solution Approach 1:
The patent extracts only the essential semantic features from defect descriptions using natural language processing, removing redundant and noisy information. The system identifies and extracts key defect characteristics (type, location, severity) while discarding irrelevant details, thereby improving classification accuracy by working with purified feature representations.
Solution Approach 2:
The patent introduces an intermediary NLP processing layer between image feature extraction and classification. This intermediary layer processes raw defect descriptions, converts unstructured text into structured semantic features, and filters out noise before feeding information to the classification model, thereby improving accuracy while eliminating redundant information.
2Adaptability or versatility
If semi-supervised learning algorithms are used for defect detection, then known defect classes can be handled, but the system fails to dynamically expand to accommodate new labels
Solution Approach 1:
The patent implements a dynamic defect classification system where new defect labels can be added without retraining the entire model. The system dynamically adapts to new defect types by learning from少量 labeled examples and integrating them into the existing classification framework, enabling continuous expansion of defect detection capabilities.
Solution Approach 2:
The patent creates a universal defect detection framework that handles both known and new defect classes through a unified architecture. The system uses transfer learning and feature reuse to accommodate multiple defect types without requiring separate specialized models for each defect class, thereby enabling dynamic expansion while managing complexity.
3Measurement precision
If manual labeling is performed for training datasets, then accurate defect labels can be obtained, but the process is time-consuming and labor-intensive
Solution Approach 1:
The patent implements a self-service labeling system where the defect detection model automatically generates labels for training data by analyzing defect images and descriptions. The system uses its own learned features to annotate new data, eliminating the need for manual labeling while maintaining high accuracy through self-consistent labeling based on learned defect characteristics.
Solution Approach 2:
The patent performs preliminary automated labeling of training data using the defect detection model before formal training begins. This preliminary action provides initial labels that can be quickly reviewed and corrected if needed, significantly reducing the time required for complete manual labeling while maintaining accuracy through a hybrid approach.
Data Source
AI summary
Systems and methods described herein relate to semantic feature extraction for auto-labeling of defects. An image is processed to obtain structured text data describing a target defect of an item appearing in the image. The structured text data is processed to convert the structured text data into a target embedding associated with the target defect. The target embedding is automatically compared with a plurality of reference embeddings to obtain a comparison result. The plurality of reference embeddings is stored in a reference data structure that associates each reference embedding with a respective defect label. Based on the comparison result, the reference data structure is updated and assignment of a target defect label to the image is initiated. Such operations may be performed for each of a plurality of images in an unlabeled dataset.