Deep learning model segmentation quantization method and system for multimodal feature distribution

By employing a segmented quantization strategy and the golden section search algorithm, efficient and accurate quantization is achieved for the multi-peak feature distribution of the target detection model. This solves the accuracy and hardware compatibility issues of traditional quantization methods in multi-peak distribution scenarios, ensuring the effective deployment of the model on edge computing devices.

CN121743813BActive Publication Date: 2026-06-26HUBEI UNIV OF TECH
View PDF 2 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
HUBEI UNIV OF TECH
Filing Date
2026-03-02
Publication Date
2026-06-26

AI Technical Summary

Technical Problem

When dealing with multi-peak feature distributions, existing target detection models cannot simultaneously meet the high-precision requirements of the main peak region and the effective coverage of the secondary peak region using traditional quantization methods. This leads to a significant increase in quantization error or loss of key information, making it difficult to maintain detection accuracy and hardware compatibility in low-bit quantization scenarios.

Method used

A segmented quantization strategy based on region partitioning is adopted. Through an adaptive quantization step size allocation mechanism, the feature space is divided into a dense central region and a sparse edge region. The golden section search algorithm is used to determine the optimal breakpoint and a differentiated quantization step size is allocated for segmented linear quantization.

Benefits of technology

It effectively reduces quantization error, maintains the detection accuracy of the model in complex open-world scenarios, and achieves efficient inference on hardware, solving the problems of computational complexity and accuracy loss of traditional quantization methods in multi-peak distribution scenarios.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN121743813B_ABST
    Figure CN121743813B_ABST
Patent Text Reader

Abstract

The application provides a deep learning model segmentation quantization method and system for multimodal feature distribution, the method comprising: collecting a feature tensor to be quantized in a quantization parameter, and determining a feature quantization range; constructing a quantization error function based on the feature quantization range, minimizing the quantization error function based on a golden section search algorithm, and determining an optimal breakpoint; dividing the feature region according to the optimal breakpoint and setting a quantization step, and constructing a segmented quantization operator; implementing segmented symmetric linear quantization according to the feature quantization range, determining the region to which the floating point value belongs through the segmented quantization operator, and performing differential mapping based on the region; and performing multiplication and addition operation on the quantized integer and the quantized weight to obtain an integer result and output. The application proposes a segmented quantization strategy based on region division, aiming to realize accurate quantization of multimodal feature distribution through an adaptive quantization step allocation mechanism.
Need to check novelty before this filing date? Find Prior Art

Citation Information

Patent Citations

  • High-precision quantification method for non-uniformly distributed data

    CN120354065A

  • Piecewise quantization method for artificial neural networks

    EP3816874A2