Scholarlink

<p>In this study, we verify whether computer-generated imagery (CGI) for real image domain adaptation using a generative adversarial network (GAN) based method can improve the accuracy of wood ear mushroom detection. Automated harvesting is important to address labor shortages in agriculture. However, a significant amount of training data is needed to detect crops with features such as wood ear mushroom. In our previous work, a slight improvement in accuracy was achieved with a small number of real-world images by incorporating CGI into the training data. Nevertheless, the reality gap between the CGI and real-world images is expected to degrade detection accuracy. GAN-based methods are effective in bridging the gap, and this study validates this approach. As a result of the experiment, the accuracy improved by approximately 6% in the F2-score between the CGI and GAN-generated datasets. However, there was no significant difference in detection accuracy when training with a combination of real-world and generated image data. In conclusion, this experiment suggests that data augmentation with generated images is not effective for this specific detection target. Nevertheless, this approach should be explored further by generating data on a larger scale with greater computational resources.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

The effectiveness of GAN-based synthetic-to-real domain adaptation methods in training a wood ear mushroom detection model

  • 期刊论文

  • 发表日期：

    2025年1月4日

查看全文
  • Hajime Taguchi,
  • Ryo Matsumura,
  • Hironori Kitakaze

摘要

In this study, we verify whether computer-generated imagery (CGI) for real image domain adaptation using a generative adversarial network (GAN) based method can improve the accuracy of wood ear mushroom detection. Automated harvesting is important to address labor shortages in agriculture. However, a significant amount of training data is needed to detect crops with features such as wood ear mushroom. In our previous work, a slight improvement in accuracy was achieved with a small number of real-world images by incorporating CGI into the training data. Nevertheless, the reality gap between the CGI and real-world images is expected to degrade detection accuracy. GAN-based methods are effective in bridging the gap, and this study validates this approach. As a result of the experiment, the accuracy improved by approximately 6% in the F2-score between the CGI and GAN-generated datasets. However, there was no significant difference in detection accuracy when training with a combination of real-world and generated image data. In conclusion, this experiment suggests that data augmentation with generated images is not effective for this specific detection target. Nevertheless, this approach should be explored further by generating data on a larger scale with greater computational resources.