基于顺序分级与数据偏置修正优化的二阶段超分辨率量化方案

doi:10.19678/j.issn.1000-3428.0253191

计算机工程 ›› 2026, Vol. 52 ›› Issue (6): 68-79. doi: 10.19678/j.issn.1000-3428.0253191

基于顺序分级与数据偏置修正优化的二阶段超分辨率量化方案

郝亮¹^,², 苏博何俊¹, 王京华¹, 徐勇¹^,³^,*()

1. 哈尔滨工业大学(深圳)信息学部, 广东深圳 518055
2. 河钢数字技术股份有限公司, 河北石家庄 050035
3. 深圳市视觉目标检测与判识重点实验室, 广东深圳 518055

收稿日期:2025-10-23 修回日期:2026-02-10 出版日期:2026-06-15 发布日期:2026-03-27
通讯作者: 徐勇
作者简介:
郝亮, 男, 正高级工程师、博士研究生, 主研方向为工业机器视觉
苏博何俊, 硕士研究生
王京华, 副教授
徐勇(通信作者), 教授
基金资助:
河钢集团重点科技项目(HG2025129); 深圳市技术攻关项目(JSGG20220831104402004)

A Two-Phase Super-Resolution Quantization Scheme Optimized by Sequential Grading and Data Bias Correction

HAO Liang¹^,², SU Bohejun¹, WANG Jinghua¹, XU Yong¹^,³^,*()

1. Division of Information Science, Harbin Institute of Technology, Shenzhen, Shenzhen 518055, Guangdong, China
2. HBIS Digital Technology Co., Ltd., Shijiazhuang 050035, Hebei, China
3. Shenzhen Key Laboratory of Visual Object Detection and Recognition, Shenzhen 518055, Guangdong, China

Received:2025-10-23 Revised:2026-02-10 Online:2026-06-15 Published:2026-03-27
Contact: XU Yong

摘要/Abstract

摘要：

模型量化技术通过将高精度浮点数映射到低比特离散空间, 能有效降低模型存储与计算开销。如何合理化地考虑参数分布的特点以构建更好的映射方案是模型量化研究的核心。现有后训练量化(PTQ)方案多默认非激活层数据呈对称钟型分布, 却忽略了激活层与输入引入的微小偏置可能导致的分布不对称性, 进而使量化映射偏向一侧, 引入显著近似损失。本文针对图像超分辨率(SR)量化方案进行研究, 对性能获得公认的二阶段PTQ方案进行改进。将预寻找量化上下界过程中的基于最值的均等分配改进为基于排序的非均匀分配, 并在伪量化过程中引入偏置量, 将数据的一部分及均值进行自适应偏移, 缓解数据偏置引发的估计损失。改进后的方案在效果上全面超越原方案, 并同样具有高压缩比以及加速比: 相比原SwinIR-light模型参数量减少60%以上, 超分辨率过程加速3倍以上。

关键词: 模型量化, 图像超分辨率, 数据偏移, 后训练量化, 排序, 偏置

Abstract:

Model quantization technology effectively reduces model storage and computational overhead by mapping high-precision floating-point data to low-bit discrete spaces. A core focus of model quantization research is how to rationally account for the characteristics of parameter distributions to construct superior mapping schemes. Existing Post-Training Quantization (PTQ) schemes nearly universally assume that the data distribution of non-activation layers follows a symmetric bell-shaped curve, but overlook the fact that small biases introduced by the model's activation layers and inputs induce distributional asymmetry. Consequently, the resulting quantization mapping is skewed to one side due to this subtle asymmetry, leading to significant approximation loss. This paper investigates quantization schemes for image Super-Resolution (SR) and proposes improvements to the widely recognized two-stage PTQ scheme. First, the max-min-based equal partitioning employed in the pre-search for quantization bounds is modified to a sorting-based non-uniform partitioning approach. Second, a bias term is introduced during the pseudo-quantization process, where a portion of the data and its mean are adaptively shifted to mitigate estimation loss caused by data bias. The improved scheme outperforms the original counterpart across all performance metrics while maintaining comparable high compression ratio and acceleration ratio—compared to the original SwinIR-light model, it reduces the parameter count by more than 60% and accelerates the SR process by more than 3 times.

Key words: model quantization, image Super-Resolution (SR), data bias, Post-Training Quantization (PTQ), sorting, bias

郝亮, 苏博何俊, 王京华, 徐勇. 基于顺序分级与数据偏置修正优化的二阶段超分辨率量化方案[J]. 计算机工程, 2026, 52(6): 68-79.

HAO Liang, SU Bohejun, WANG Jinghua, XU Yong. A Two-Phase Super-Resolution Quantization Scheme Optimized by Sequential Grading and Data Bias Correction[J]. Computer Engineering, 2026, 52(6): 68-79.

https://www.ecice06.com/CN/Y2026/V52/I6/68

图/表 12

图1 标准正态分布与偏移正态分布

Fig.1 Standard normal distribution and the biased normal distribution

图2 Y(激活后的X)、Z、YZ的频率分布直方图

Fig.2 The frequency distribution histograms of Y (activated X), Z, and YZ

图3 2DQuant+二阶段量化方案的量化过程示意图

Fig.3 Schematic diagram of the quantization process for the two-stage quantization scheme 2DQuant+

图4 5个标准数据集图像像素灰度值的频率分布直方图

Fig.4 Histograms of frequency distribution of image pixel gray values across five standard datasets

图5 使用BC操作补偿均值偏置

Fig.5 Compensating for the mean bias using the BC operation

图6 2DQuant+方案与原2DQuant方案在img093图像上的超分辨率效果

Fig.6 Super-resolution effects between the 2DQuant+ scheme and the original 2DQuant scheme on img093 image

图7 2DQuant+方案与原2DQuant方案在img040图像上的超分辨率效果

Fig.7 Super-resolution effects between the 2DQuant+ scheme and the original 2DQuant scheme on img040 image

参考文献 26

1	DONG C, LOY C C, HE K, et al. Image super-resolution using deep convolutional networks[EB/OL]. [2025-09-30]. https://arxiv.org/abs/1501.00092.
2	KIM J, LEE J K, LEE K M. Accurate image super-resolution using very deep convolutional networks[EB/OL]. [2025-09-30]. https://arxiv.org/abs/1511.04587.
3	WANG X T, YU K, WU S X, et al. ESRGAN: enhanced super-resolution generative adversarial networks[EB/OL]. [2025-09-30]. https://arxiv.org/abs/1809.00219.
4	ZHANG Y L, LI K P, LI K, et al. Image super-resolution using very deep residual channel attention networks[EB/OL]. [2025-09-30]. https://arxiv.org/abs/1807.02758.
5	VASWANI A, SHAZEER N, PARMAR N, et al. Attention is all you need[EB/OL]. [2025-09-30]. https://arxiv.org/abs/1706.03762.
6	LIU K, QIN H, GUO Y, et al. 2DQuant: low-bit post-training quantization for image super-resolution[C]//Proceedings of the 38th Conference on Neural Information Processing Systems (NeurIPS 2024). New York, USA: Curran Associates Inc., 2024: 71068-71084.
7	ZHONG Y S, LIN M B, XIE J J, et al. Distribution-flexible subset quantization for post-quantizing super-resolution networks[EB/OL]. [2025-09-30]. https://arxiv.org/abs/2305.05888.
8	LI H X, YAN C Q, LIN S H, et al. PAMS: quantized super-resolution via parameterized max scale[C]//Proceedings of the European Conference on Computer Vision (ECCV). Berlin, Germany: Springer, 2020: 564-580.
9	QIN H T, ZHANG Y L, DING Y F, et al. QuantSR: accurate low-bit quantization for efficient image super-resolution[C]//Advances in Neural Information Processing Systems. New York, USA: Curran Associates, Inc., 2023: 56838-56848.
10	LIANG J Y, CAO J Z, SUN G L, et al. SwinIR: image restoration using swin transformer[EB/OL]. [2025-09-30]. https://arxiv.org/abs/2108.10257.
11	HENDRYCKS D, GIMPEL K. Gaussian Error Linear Units (GELUs)[EB/OL]. [2025-09-30]. https://arxiv.org/abs/1606.08415.
12	LIM B, SON S, KIM H, et al. Enhanced deep residual networks for single image super-resolution[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Honolulu, USA: IEEE Press, 2017: 1132-1140.
13	TIMOFTE R, AGUSTSSON E, VAN GOOL L, et al. NTIRE 2017 challenge on single image super-resolution: methods and results[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW). Honolulu, USA: IEEE Press, 2017: 1110-1121.
14	李锦辉, 刘继, 闵兰. 基于领域知识蒸馏的旅游文本情感分类轻量模型. 重庆邮电大学学报(自然科学版), 2025, 37 (4): 617- 626.
	LI J H , LIU J , MIN L . Lightweight sentiment classification model for tourism texts based on domain knowledge distillation. Journal of Chongqing University of Posts and Telecommunications (Natural Science Edition), 2025, 37 (4): 617- 626.
15	HINTON G, VINYALS O, DEAN J. Distilling the knowledge in a neural network[EB/OL]. [2025-09-30]. https://arxiv.org/abs/1503.02531.
16	BEVILACQUA M, ROUMY A, GUILLEMOT C, et al. Low-complexity single-image super-resolution based on nonnegative neighbor embedding[C]//Proceedings of the British Machine Vision Conference 2012. Surrey, UK: British Machine Vision Association, 2012: 1-10.
17	ZEYDE R, ELAD M, PROTTER M. On single image scale-up using sparse-representations[C]//Proceedings of the International Conference on Curves and Surfaces (ICCS). Berlin, Germany: Springer, 2010: 711-730.
18	MARTIN D, FOWLKES C, TAL D, et al. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics[C]//Proceedings of the 8th IEEE International Conference on Computer Vision. Vancouver, Canada: IEEE Press, 2002: 416-423.
19	HUANG J B, SINGH A, AHUJA N. Single image super-resolution from transformed self-exemplars[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Boston, USA: IEEE Press, 2015: 5197-5206.
20	MATSUI Y , ITO K , ARAMAKI Y , et al. Sketch-based manga retrieval using Manga109 dataset. Multimedia Tools and Applications, 2017, 76 (20): 21811- 21838.
21	WU G , JIANG J J , JIANG K , et al. Fully 1×1 convolutional network for lightweight image super-resolution. Machine Intelligence Research, 2024, 21 (6): 1062- 1076.
22	JACOB B, KLIGYS S, CHEN B, et al. Quantization and training of neural networks for efficient integer-arithmetic-only inference[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Salt Lake City, USA: IEEE Press, 2018: 2704-2713.
23	LI R D, WANG Y, LIANG F, et al. Fully quantized network for object detection[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Long Beach, USA: IEEE Press, 2019: 2805-2814.
24	TU Z J, HU J, CHEN H T, et al. Toward accurate post-training quantization for image super resolution[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Vancouver, Canada: IEEE Press, 2023: 5856-5865.
25	王浩, 端木春江. 极度轻量化的实时4K图像超分辨率重建网络. 微电子学与计算机, 2025, 42 (5): 73- 80.
	WANG H , DUANMU C J . Extremely lightweight real-time 4K image super-resolution network. Microelectronics & Computer, 2025, 42 (5): 73- 80.
26	王拓然, 程娜, 丁士佳, 等. 基于自适应注意力融合特征提取网络的图像超分辨率. 计算机应用研究, 2023, 40 (11): 3472-3477, 3508.
	WANG T R , CHENG N , DING S J , et al. Image super-resolution based on adaptive attention fusion feature extraction network. Application Research of Computers, 2023, 40 (11): 3472-3477, 3508.

[1]	张俊娜, 王泓尊, 丁春涛. 后训练量化方法综述(特邀)[J]. 计算机工程, 2026, 52(1): 33-60.
[2]	魏铭康, 李嘉楠, 韩林, 高伟, 赵荣彩, 王洪生. 面向深度学习编译器的多粒度量化框架支持与优化[J]. 计算机工程, 2025, 51(5): 62-72.
[3]	赵瑶谦, 滕奇志, 何小海, 税爱, 陈洪刚. 基于自注意力特征蒸馏的轻量级图像超分辨率重建[J]. 计算机工程, 2025, 51(5): 257-265.
[4]	裴欣, 佟继周, 马福利, 于勤思, 杨广建. ASO-S卫星源包数据拼接与排序策略研究[J]. 计算机工程, 2025, 51(5): 166-176.
[5]	郭俊辰, 马御棠, 相艳, 赵学东, 郭军军. 基于Prompt打分的实体链接方法[J]. 计算机工程, 2025, 51(3): 334-341.
[6]	王月昊, 周若华. 低资源环境下的语音唤醒研究综述[J]. 计算机工程, 2025, 51(2): 35-53.
[7]	阳佩珉, 闵华松. 嵌入式肌电腕带实时采集与识别系统设计[J]. 计算机工程, 2025, 51(2): 259-268.
[8]	李伟康, 张思全. 掩模特征融合: 实例分割新范式[J]. 计算机工程, 2025, 51(2): 126-138.
[9]	杨嘉卉, 尤再进, 倪立夫, 赵煜, 李婉莹. 集装箱码头泊位-岸桥减排协同调度优化研究[J]. 计算机工程, 2025, 51(10): 381-391.
[10]	钟忺, 陈亮, 刘文璇, 叶舒, 江奎, 王正, 林嘉文. 时空语义驱动的渐进多视角行为去偏置研究[J]. 计算机工程, 2025, 51(1): 1-10.
[11]	张国富, 管燕妮, 苏兆品, 岳峰. 多阶段应急物资多目标连续分配问题建模与求解[J]. 计算机工程, 2024, 50(12): 329-345.
[12]	王磊, 李文杰, 王海. 改进概率语言ORESTE的人岗匹配决策方法[J]. 计算机工程, 2024, 50(12): 407-416.
[13]	李志鹏, 陈丹阳, 钟诚. 一种改进的超分辨率轻量化特征融合方法[J]. 计算机工程, 2024, 50(11): 258-265.
[14]	孙瑾, 苏文娟, 王璐, 叶克鑫. 基于联盟区块链和星际文件系统的安全租房方案[J]. 计算机工程, 2024, 50(11): 187-196.
[15]	乔艺萌, 荆一楠, 张寒冰. 健壮且自适应的学习型近似查询处理方法研究[J]. 计算机工程, 2024, 50(1): 30-38.

选择文件类型/文献管理软件名称

选择包含的内容

基于顺序分级与数据偏置修正优化的二阶段超分辨率量化方案

A Two-Phase Super-Resolution Quantization Scheme Optimized by Sequential Grading and Data Bias Correction

RichHTML

PDF

可视化

被引次数

摘要/Abstract

引用本文

使用本文

图/表 12

参考文献 26

相关文章 15

编辑推荐

Metrics

本文评价

模态框（Modal）标题

选择文件类型/文献管理软件名称

选择包含的内容

基于顺序分级与数据偏置修正优化的二阶段超分辨率量化方案

A Two-Phase Super-Resolution Quantization Scheme Optimized by Sequential Grading and Data Bias Correction

RichHTML

PDF

可视化

被引次数

摘要/Abstract

引用本文

使用本文

图/表 12

参考文献 26

相关文章 15

编辑推荐

Metrics

本文评价