An image recognition model pruning and compression method and system based on intra-class response evaluation driving

By adopting an image recognition model pruning and compression method based on intra-class response evaluation, the problem of poor model stability under noisy label conditions in the existing technology is solved, and efficient pruning and compression in noisy environments is achieved, maintaining model performance and reducing the number of parameters.

CN121981186BActive Publication Date: 2026-06-26NANJING UNIV OF SCI & TECH
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202610435584.8
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2026-04-03
Publication Date
2026-06-26
Estimated Expiration
2046-04-03

AI Technical Summary

Technical Problem

Most existing image recognition model pruning and compression methods are based on the assumption of clean and accurately annotated data, which leads to label noise in real-world scenarios, reduces the performance of deep neural networks, and shows high sensitivity to noisy label conditions, resulting in poor stability. Furthermore, they ignore the uniqueness of the data and have poor compression effects.

Method used

A method based on intra-class response evaluation is adopted to select clean and noisy image samples through JS divergence, calculate the importance score of filter parameters, and perform cross-class weighted aggregation to achieve pruning and compression of the image recognition model, retaining clean data information and eliminating the influence of noise.

Benefits of technology

While maintaining the performance of the image recognition model, the parameters of the deep neural network are reduced to improve the efficiency and effect of pruning and compression, thereby enhancing the stability of the model.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN121981186B_ABST
    Figure CN121981186B_ABST
Patent Text Reader

Abstract

The application discloses a kind of based on in-class response evaluation drive's image recognition model pruning compression method and system, first, initial noisy image dataset is selected using JS divergence and respectively obtains clean image sample and noise image sample, then clean image sample and noise image sample are classified and clean sample subset and noise sample subset are obtained;The application realizes the function that retains the parameter mainly captured from clean data information and eliminates the parameter mainly affected by noise data, and by classifying the importance of each subset parameter and using the weighted aggregation strategy to aggregate the importance score, the importance of the parameter can be accurately calculated according to the response of the parameter to clean data and noise data, which can not only reduce the deep neural network parameter while maintaining the performance of image recognition model, but also improve the pruning compression efficiency and effect of image recognition model, and is suitable for being widely promoted and used.
Need to check novelty before this filing date? Find Prior Art

Citation Information

Patent Citations

  • Image compressed sensing reconstruction method, system, equipment and medium

    CN114742911A

  • Noise-containing label image identification method and system based on category balance sample selection

    CN117173494A