Rule supervision-based image recognition verification learning method and device

By proposing a rule-based supervised image recognition verification learning method, this paper addresses the problems of insufficient symbolic information and solution space expansion in deep learning models without data annotation. It achieves efficient and low-cost unsupervised image recognition, applicable to various rule-driven tasks, and improves the computational efficiency and recognition accuracy of the model.

CN122135110APending Publication Date: 2026-06-02NANJING UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202610336373.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-03-19
Publication Date
2026-06-02

AI Technical Summary

Technical Problem

Existing technologies, when lacking data annotation, suffer from problems such as insufficient symbolic information, solution space expansion, and rampant shortcuts in deep learning image recognition models. These issues result in weak technical versatility, low computational efficiency, and difficulty in effective application in unsupervised scenarios.

Method used

A rule-supervised image recognition verification learning method is adopted. By acquiring image datasets and rule bases, neural networks are used to identify symbol information, establish a score index stack structure, dynamically generate symbol combinations and perform rule verification, and adjust the model output by combining prior distribution to optimize neural network parameters.

Benefits of technology

It significantly reduces data annotation costs, improves computational efficiency and technical versatility in unsupervised scenarios, adapts to various rule-driven tasks, and enhances the reliability and interpretability of image recognition results.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122135110A_ABST
    Figure CN122135110A_ABST
Patent Text Reader

Abstract

This invention discloses a rule-supervised image recognition verification learning method and apparatus, comprising: a symbol recognition stage, in which an image set is input into a neural network model to obtain a set of symbol probabilities predicted by the model, a score index for evaluating the quality of symbol combinations is determined, and a heap structure is established based on the score index according to a dynamic combination ranking algorithm; and a dynamic verification stage, in which a knowledge verification function V is used to verify the symbol combination. KB The heap structure is dynamically maintained by sequentially determining whether the top symbol combination satisfies the knowledge rule based on the order of symbol combinations in the heap structure, until the first verified combination is found, thus finding the combination that satisfies V. KB The optimal symbolic combination solution with "constraint and maximum score index" is found; feedback optimization stage: the model output distribution is adjusted according to the prior distribution information to alleviate the inference shortcut problem, and the found optimal combination solution is used as supervision information to feed back to the model for gradient optimization, thus completing the image recognition training with rules as supervision information.
Need to check novelty before this filing date? Find Prior Art