基于改进生成对抗网络的甜樱桃数据增强方法
作者:
作者单位:

作者简介:

通讯作者:

中图分类号:

基金项目:

山东省现代农业产业技术体系果品产业创新团队项目(SDAIT-06-12)


Data Augmentation Method for Sweet Cherries Based on Improved Generative Adversarial Network
Author:
Affiliation:

Fund Project:

  • 摘要
  • |
  • 图/表
  • |
  • 访问统计
  • |
  • 参考文献
  • |
  • 相似文献
  • |
  • 引证文献
  • |
  • 资源附件
  • |
  • 文章评论
    摘要:

    为解决在数据不平衡条件下甜樱桃分类模型出现的长尾类不平衡问题,提出了一种基于深度卷积生成对抗网络(Deep convolutional generative adversarial networks,DCGAN)的缺陷甜樱桃图像增强方法。首先,在生成器部分引入多尺度残差块(MSRB)和CBAM注意力机制,增强了模型特征表达能力和生成图像细节质量,同时改善了梯度流;在判别器部分应用谱归一化技术,并引入Wasserstein距离和加梯度惩罚的损失函数,增强了模型训练稳定性和收敛速度。实验结果表明,与传统的GAN模型相比,本文模型可以生成更高质量的缺陷甜樱桃图像,两种缺陷甜樱桃图像的FID值(Fréchet inception distance)分别为64.36和59.97。本文模型生成的数据增强后,VGG19和MobileNetV3的甜樱桃分类准确率分别提高16.44个百分点和13.94个百分点。

    Abstract:

    To address the class imbalance in sweet cherry data, a novel image enhancement method based on sweet cherry generative adversarial network, SCGAN was proposed. The generator incorporated multi-scale residual blocks (MSRB) and the convolutional block attention module (CBAM), enhancing the model’s feature representation and the quality of generated images. These blocks captured features at various scales, and CBAM focused on channel and spatial information, improving image quality. In the discriminator, spectral normalization and the Wasserstein distance with a gradient penalty loss function were applied. This combination controled the discriminator’s power, prevented overfitting, and boosted training stability and speed. Experimental results showed that SCGAN produced higher quality defective sweet cherry images compared with traditional GANs, with Fréchet inception distance (FID) scores of 64.36 and 59.97 for two types of defects. After data augmentation with SCGAN, classification accuracy for VGG19 and MobileNetV3 was increased by 16.44 percentage points and 13.94 percentage points, respectively. The data augmentation method presented held significant potential in addressing data imbalance issues within the agricultural and food sectors. It not only improved the generalization capability of models but also provided a more reliable data foundation for practical applications. Through this approach, it was possible to more effectively tackle long-tail class imbalance issues, which enhanced the accuracy and efficiency of agricultural and food detection systems.

    参考文献
    相似文献
    引证文献
引用本文

韩翔,李玉强,高昂,马静怡,宫庆福,宋月鹏.基于改进生成对抗网络的甜樱桃数据增强方法[J].农业机械学报,2024,55(10):252-262. HAN Xiang, LI Yuqiang, GAO Ang, MA Jingyi, GONG Qingfu, SONG Yuepeng. Data Augmentation Method for Sweet Cherries Based on Improved Generative Adversarial Network[J]. Transactions of the Chinese Society for Agricultural Machinery,2024,55(10):252-262.

复制
分享
文章指标
  • 点击次数:
  • 下载次数:
  • HTML阅读次数:
  • 引用次数:
历史
  • 收稿日期:2024-06-04
  • 最后修改日期:
  • 录用日期:
  • 在线发布日期: 2024-10-10
  • 出版日期:
文章二维码