| 摘要: |
| 目的 针对苹果病害中比较常见的症状———花叶病,尤其在昼夜温差大的条件下发病迅速,落叶率提高,造
成苹果大面积减产,产生巨大的经济损失;对于花叶病病斑数量太多,尺度不一的影响,从而造成病害识别准确率
较低等问题,提出了一种引入迁移学习和胶囊网络的方法,以提高病害识别率。 方法 首先对获得的花叶病数据集
进行扩充、数据增强等处理,并利用 Labelme 工具对图像进行标注,分别标记出病斑区域和叶片区域;其次将训练
好的 VGG16 模型权重通过迁移学习技术移至 U-net 中编码部分,并引入胶囊网络,使得整个网络具有更强的特征
提取能力;然后对 VGG16 模型、胶囊网络部分进行训练,最后将训练好的网络模型进行语义分割并输出测试的结
果。 结果 实验结果 表 明, 原 始 数 据 集 的 准 确 率 为 87. 51%, 引 入 迁 移 学 习 后 的 准 确 率 提 升 至 91. 78%, 提 升 了
4. 88%;引入胶囊网络的准确率提升至 90. 04%,提升了 2. 89%;而引入迁移学习和胶囊网络之后,准确率提升至
93. 42%,提升了 6. 75%。 并且模型每一轮的训练时间也在引入了迁移学习后提升了 2 s。 结论 据实验结果可以证
明模型方法引入迁移学习和胶囊网络后,相较于传统模型在识别准确率方面有了一定的提升,其次也减少了每一
轮的模型训练时间,总体分割性能较好。 |
| 关键词: 病害识别 花叶病 病斑 VGG16 U-net 胶囊网络 |
| DOI： |
| 分类号: |
| 基金项目: |
|
| Research on Image Semantic Segmentation Structure Based on U-net and Capsule Network |
|
LIU Xiangju, ZHAO Huimeng, FANG Xianjin
|
|
School of Computer Science and Engineering Anhui University of Science & Technology Anhui Huainan 232001 China
|
| Abstract: |
| Objective Mosaic disease is a common symptom in apples. Especially under the condition of large temperature
differences between day and night the onset of mosaic disease is rapid which can lead to an increase in the rate of
defoliation resulting in a large reduction in apple production and huge economic losses. The number of mosaic disease
spots is too high and the scale of the spots varies resulting in problems such as low accuracy of disease identification.
Based on this a method that introduces transfer learning and capsule networks was proposed to improve the disease
identification rate. Methods Firstly this method expanded and enhanced the obtained mosaic disease dataset and used
the Labelme tool to annotate images marking the lesion area and leaf area respectively. Secondly the weight of the
trained VGG16 model was transferred to the coding part of U-net through the transfer learning technology and the capsule
network was introduced so that the whole network had a stronger feature extraction ability. Then the VGG16 model and
capsule network were trained. Finally the trained network model was semantically segmented and test results were
output. Results The experimental results showed that the accuracy rate of the original dataset was 87. 51% and the
accuracy rate after the introduction of transfer learning was improved to 91. 78% an improvement of 4. 88% the accuracy
of introducing the capsule network was improved to 90. 04% an increase of 2. 89% after the introduction of transfer learning and capsule network the accuracy rate was improved to 93. 42% an improvement of 6. 75%. In addition the
training time of each round of the model was also improved by 2 s after the introduction of transfer learning.
Conclusion According to the experimental results it can be proved that the proposed model after the introduction of
transfer learning and capsule network has a certain improvement in identification accuracy compared with the traditional
model. Furthermore this method also reduces the model training time in each round and the overall segmentation
performance is better. |
| Key words: disease identification mosaic disease disease spots VGG16 U-net capsule network |
