A training and labeling parallel collaboration method and device based on defect information query

CN116796028BActive Publication Date: 2026-07-07XIAMEN WEIYA INTELLIGENT TECHNOLOGY CO LTD +1
View PDF 1 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
XIAMEN WEIYA INTELLIGENT TECHNOLOGY CO LTD
Filing Date
2023-05-23
Publication Date
2026-07-07

AI Technical Summary

Technical Problem

In the field of industrial image detection, existing deep learning methods rely on a large amount of expert experience annotation, which leads to high data acquisition and annotation costs and unbalanced data distribution, making it difficult to effectively apply to deep learning models.

Method used

A parallel collaborative training and labeling method based on defect information query is adopted. By acquiring a small amount of labeled dataset and a large amount of unlabeled dataset, a hybrid query strategy of coarse screening and fine screening is combined. Feature extraction and sample screening are performed using backbone network and query network, and a deep detection model is trained in parallel with selective labeling.

Benefits of technology

It reduces data annotation costs and time costs, quickly filters valuable data, improves the model's detection performance and generalization ability, avoids overfitting, and achieves efficient deep detection model construction.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN116796028B_ABST
    Figure CN116796028B_ABST
Patent Text Reader

Abstract

The application discloses a training and labeling parallel cooperation method, device and equipment, which comprises the following steps: obtaining a small amount of labeled data set and a large amount of unlabeled data set; taking the labeled data set as the training data set for model initialization, training to obtain a detection model and a query network; performing sample coarse screening on the unlabeled data set through a batch query module to obtain multiple batches of data samples; inputting the first batch of samples in the multiple batches of data samples into the detection model to extract feature information and then inputting the feature information into the query network to perform sample fine screening, thereby obtaining K samples; inputting the K samples into the training data set after labeling to update the detection model; continuing to screen the second batch of samples in the multiple batches of data samples for next iteration until the detection performance of the detection model or the labeling cost upper limit is reached. The valuable data can be quickly screened, the model can be efficiently constructed, and the demand performance of the business can be met.
Need to check novelty before this filing date? Find Prior Art