Quantization training method and device, equipment and computer readable storage medium

By expanding the channels and distilling intermediate supervision knowledge in the floating-point network model, the quantization training process is optimized, the noise problem of the quantization model in pixel-level tasks is solved, and the performance of the quantization model and the image processing effect are improved.

CN116306820BActive Publication Date: 2026-07-24伟光有限公司(CN)
View PDF 2 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
伟光有限公司(CN)
Filing Date
2023-03-24
Publication Date
2026-07-24

AI Technical Summary

Technical Problem

Existing Quantization-Aware Training (QAT) algorithms have significant limitations in pixel-level tasks such as image denoising, deblurring, and super-resolution. Compared to the original model, the quantized model exhibits noticeable noise in local image regions that is easily distinguishable to the naked eye, resulting in low performance.

Method used

By expanding the channels of the floating-point network model, calculating feature similarity to determine the distillation loss function, supervising the quantization-aware training process, introducing knowledge distillation of intermediate supervision mode, optimizing the quantization training process, and reducing the model performance loss caused by quantization error.

Benefits of technology

Without changing the model deployment, the performance of the quantization model was improved, the accuracy loss caused by quantization error was reduced, and the accuracy of image processing was enhanced.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN116306820B_ABST
    Figure CN116306820B_ABST
Patent Text Reader

Abstract

Embodiments of the present application disclose a kind of quantization training method, device and computer readable storage medium.The method comprises: based on the sub-feature of each sub-network in floating point network model, channel expansion is carried out, the first sub-feature of each first sub-network in first network model is determined;Based on each first sub-feature and each initial second sub-feature of each second sub-network in second network model, the similarity of feature is calculated, and the distillation loss function is determined;Based on distillation loss function, the quantization perception training process of second network model is supervised, and the quantization parameter is determined;Quantization parameter is used in the process of using second network model, and the network parameter of second network model is quantized to determine the quantization model used for image processing.
Need to check novelty before this filing date? Find Prior Art