首页|一种用于智能零售视觉结算的增量学习方法

一种用于智能零售视觉结算的增量学习方法

扫码查看
原文链接
  • 国家科技期刊平台
  • NETL
  • NSTL
  • 万方数据
为了解决智能零售视觉结算任务中的增量学习问题，提出一个新颖的基于数据增广的三元组模型，由合成网络、渲染网络和检测网络组成。合成网络与渲染网络协同学习，将单品示例图像以数据增广方式生成分布接近真实数据的渲染视觉结算图像。在增量学习阶段，原有产品和新产品的示例图像协同学习，生成包含新产品的结算图像。所有渲染视觉结算图像被用于训练检测网络，训练好的模型能同时识别原有产品和新产品。试验结果表明，该模型具有相较于现有增量学习方法更优异的对抗灾难性遗忘能力，增量后的结算准确率为 64。90%，遗忘率为3。63%，优于现有最佳方法的4。38%。
Incremental learning method for intelligent retail automatic check-out
To deal with the incremental learning issue of intelligent retail automatic check-out,a novelty data-argument-based triplet model is proposed,which consists of the synthesizer network,the renderer network and the detector network.Specifically,the synthesizer network and the renderer network learn collaboratively to generate rendered check-out images with distribution close to the real data by synthesizing and rendering the single-product example images with data augmentation.In the incremental learning phase,the original and new product example images are collaborative learned to generate rendered check-out images containing new products.These rendered check-out images are utilized to train the product detector network.The model obtained by training in this way has the ability to recognize both of original products and new products.The experimental results show that the model has more excellent ability to overcome catastrophic forgetting compared to the existing incremental learning methods.The incremented check-out accuracy is 64.90% with a forgetting rate of 3.63% ,which is better than the state-of-the-art method of 4.38% .

automatic check-outincremental learningsynthesizerrendererobject detection

陈昊、魏秀参、肖亮

展开 >

南京理工大学 计算机科学与工程学院,江苏 南京 210094

视觉结算 增量学习 合成 渲染 目标检测

国家重点研发计划青年科学家项目国家自然科学基金国家自然科学基金江苏省自然科学基金青年基金中国人工智能学会-华为MindSpore学术奖励基金江苏省地质局科研项目

2021YFA10011006227223161871226BK20210340CAAIXSJLJJ-2022-001B2023KY11

2024

10.14177/j.cnki.32-1397n.2024.48.01.007

南京理工大学学报(自然科学版)
南京理工大学

南京理工大学学报(自然科学版)

CSTPCD北大核心
影响因子：0.526
ISSN：1005-9830
年,卷(期)：2024.48(1)
  • 26