| 摘要: |
| 目的 为了提高果蔬农产品识别的准确性，使果蔬农产品分类实现自动化。方法 利用深度卷积神经网路强大的特征学习和特征表达能力，来自动学习果蔬种类特征，提出基于位置的柔性注意力算法，对Inceptionv3神经网络进行改进，并结合参数迁移学习方法建立果蔬识别模型；针对果蔬种类繁多，且国内外缺乏完善的果蔬图像数据库这一现状，构建果蔬图像数据集；在此数据集上将文中所提出的果蔬识别算法与其他果蔬识别算法进行对比。结果 试验结果表明，在学习率为0.1、迭代次数为5000时，文中提出算法的准确率高达97.89%。结论 相较于现有果蔬识别算法，所提出的果蔬识别算法的识别性能最优，鲁棒性最强。 |
| 关键词: 注意力机制 果蔬图像 迁移学习 图像识别 |
| DOI：10.19554/j.cnki.1001-3563.2019.21.005 |
| 分类号:TP391 |
| 基金项目:国家自然科学基金（81101116） |
|
| Fruit and Vegetable Recognition Algorithm Based on Improved Inceptionv3 |
|
JU Zhi-yong, MA Su-ping
|
|
University of Shanghai for Science and Technology, Shanghai 200093, China
|
| Abstract: |
| The work aims to improve the recognition accuracy of fruit and vegetable products, so as to automatically classify these products. The powerful feature learning and feature expression capabilities of deep convolutional neural networks were used to automatically learn the characteristics of fruit and vegetable types, the location-wise soft attention algorithm was proposed to improve the Inceptionv3 neural network, and the parameter transfer learning method was combined to establish the fruit and vegetable recognition model. In view of the wide variety of fruits and vegetables, and the lack of a complete database of fruit and vegetable images at home and abroad, the fruit and vegetable image data sets were constructed. Based on the data sets above, the proposed fruit and vegetable recognition algorithm was compared with other fruit and vegetable recognition algorithms. The experimental results showed that, when the learning rate was 0.1 and the number of iterations was 5,000, the accuracy of the proposed algorithm was as high as 97.89%. Compared with the existing fruit and vegetable recognition algorithms, the proposed fruit and vegetable recognition algorithm has the best recognition performance and the strongest robustness. |
| Key words: attention mechanism fruit and vegetable image transfer learning image recognition |
