面向小样本苗绣图像的生成与识别研究  

Research on Generation and Recognition of Few-Shot Miao Embroidery Images

作　　者：吴菁 杨邦勤 张银建 李明珠[2] 陈妍 WU Jing;YANG Bangqin;ZHANG Yinjian;LI Mingzhu;CHEN Yan(School of Data Science and Information Engineering,Guizhou Minzu University,Guiyang 550025,China;ZX-YZ School of Network Science,Haikou University of Economics,Haikou 570203,China;School of Automation Science and Engineering,South China University of Technology,Guangzhou 510641,China)

机构地区：[1]贵州民族大学数据科学与信息工程学院,贵州贵阳550025 [2]海口经济学院中芯依智网络学院,海南海口570203 [3]华南理工大学自动化科学与工程学院,广东广州510641

出　　处：《现代信息科技》2025年第2期24-32,共9页Modern Information Technology

基　　金：贵州省科技计划项目(黔科合基础-ZK[2021]一般340);贵州省教育厅自然科学研究项目(黔教技[2023]061号);贵州省教育厅自然科学研究项目(黔教技[2023]012号);贵州省教育厅自然科学研究项目(黔教技[2022]047号);贵州民族大学博士科研启动项目(GZMUZK[2024]QD11)。

摘　　要：为了解决苗绣图像样本量不足以及识别精度较低的问题,提出了一个新颖的样式生成对抗网络(StyleGAN2)结合高效通道注意力机制(ECA)优化的残差神经网络(ResNet50)苗绣图像分类模型。首先,利用StyleGAN2对原始苗绣数据集进行训练和生成,以扩充小样本数据集。接着,在ResNet50骨干网络中融入ECA注意力机制,增强了模型的特征提取能力。最终,结合StyleGAN2和ECA优化后的ResNet50创建了新的StyleGAN2-ECA-ResNet50模型,用于小样本苗绣图像的识别。实验结果显示,该方法在测试集上的准确率达到了89.29%,较传统的ResNet50模型提高了5.87%,并且在性能上超过了多个先进的深度学习模型。To address the issues of insufficient sample sizes and low recognition accuracy of Miao embroidery images,this paper proposes a novel Miao embroidery image classification model,which combines StyleGAN2 with optimized ResNet50 by the Efficient Channel Attention(ECA).Firstly,StyleGAN2 is utilized to train and generate on the original Miao embroidery dataset,thereby augmenting the Few-Shot dataset.Subsequently,the ECA Attention Mechanism is integrated into the ResNet50 backbone network to enhance the feature extraction capability of the model.Ultimately,the new StyleGAN2-ECA-ResNet50 model is created by combining StyleGAN2 and ResNet50 optimized by ECA for the recognition of Few-Shot Miao embroidery images.Experimental results show that the accuracy of this method reaches 89.29%on the test set,which is an improvement of 5.87%over the traditional ResNet50 model and surpasses several advanced Deep Learning models in performance.

关 键 词：苗绣 小样本图像分类 数据增强 StyleGAN2 ECA 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]