Method and apparatus for establishing neural network model based on dynamic mixed precision quantization, image classification method and apparatus, and electronic device

By using a dynamic mixed-precision quantization mechanism, the quantization precision of the neural network model is adjusted according to the image quality level, which solves the problem of unreasonable resource allocation in power vision scenarios, improves the computing efficiency and energy efficiency of edge devices, and meets the requirements of real-time performance and low power consumption.

CN121305462BActive Publication Date: 2026-07-24BEIJING SMARTCHIP MICROELECTRONICS TECHNOLOGY CO LTD +2
View PDF 2 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
BEIJING SMARTCHIP MICROELECTRONICS TECHNOLOGY CO LTD
Filing Date
2025-10-14
Publication Date
2026-07-24

AI Technical Summary

Technical Problem

Existing deep neural network models cannot perceive and distinguish data heterogeneity in power vision scenarios, leading to unreasonable allocation of computing resources, waste of resources or degraded recognition performance, and difficulty in meeting the requirements of real-time performance and low power consumption in edge environments.

Method used

By employing a dynamic mixed-precision quantization mechanism, the quantization precision of the neural network model is dynamically adjusted based on the classification and recognition difficulty coefficient of the image quality level. Combined with feedback optimization during the training and validation processes, computational resources can be allocated on demand.

Benefits of technology

It improves the computational efficiency and energy efficiency of the model on edge devices, ensures high robustness in recognition performance, adapts to different hardware resource budgets and environmental changes, and meets the real-time and low-power requirements of power vision tasks.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN121305462B_ABST
    Figure CN121305462B_ABST
Patent Text Reader

Abstract

The present disclosure relates to the field of artificial intelligence and computer technology, in particular to a method and device for establishing a neural network model based on dynamic mixed precision quantization, an image classification method and device, and an electronic device. The method comprises: obtaining visual monitoring image data and generating a sample set with quality levels, which is divided into a training set and a validation set; extracting feature embeddings of the training set samples based on a pre-trained model, and calculating a classification recognition difficulty coefficient for each quality level accordingly; using a preset precision quantization strategy, corresponding quantization precision is configured for network parameter configuration of the neural network model when processing samples of different quality levels according to the classification recognition difficulty coefficient, and a precision-quantized model is obtained. Thus, the model calculation efficiency is improved and the energy consumption is reduced.
Need to check novelty before this filing date? Find Prior Art