基于高光谱成像和GAN-SA-UNet算法的烟叶叶脉分割方法研究

Tobacco Leaf Vein Segmentation Method Based on Hyperspectral Imaging and GAN-SA-UNet Algorithm

摘要:

叶脉作为植物的重要特征，包含生理和遗传信息，针对复杂叶面纹理状态下的细小叶脉边缘分割模糊、分割精度低等问题，以烟叶为研究对象，提出了一种GAN-SA-UNet叶脉分割算法。通过高光谱成像技术获取叶脉与叶面光谱信息，并利用主成分分析（Principal component analysis，PCA）对其进行降维，得到合成图。在此基础上，引入空间注意力机制，捕捉关键的空间特征，提高分割精度，同时引入对抗网络，优化生成结果，提高叶脉分割的鲁棒性。结果表明：叶脉与叶面光谱PCA前3个主成分解释率达到95.71%，二者降维后的光谱特征表现出明显的可分性，前3个主成分合成图能够凸显叶面与叶脉之间的差异，突出叶脉特征。GAN-SA-UNet分割算法能够捕捉复杂叶面纹理图像的脉络特征，分割准确率和交并比分别达98.93%和66.23%，与原模型相比，分别提高0.18个百分点和4.21个百分点，单幅图像推理时间为4ms。在对不同产地、部位、等级、类型烟叶验证测试中表现出较强的泛化能力和高效准确的识别能力。

Abstract:

As an important feature of plants, leaf veins contain physiological and genetic information. Aiming at the problems of blurred edge segmentation and low segmentation accuracy of small veins in complex leaf texture state, a GAN-SA-UNet vein segmentation algorithm was proposed with tobacco leaves as the research object. The spectral information of veins and leaves was obtained by hyperspectral imaging technology, and the principal component analysis ( PCA ) was used to reduce the dimension and obtain the composite map. On this basis, the spatial attention mechanismwas introduced to capture the key spatial features and improve the segmentation accuracy. At the same time, the adversarial network was introduced to optimize the generated results and improve the robustness of vein segmentation. The results showed that the interpretation rate of the first three principal components of PCA of leaf vein and leaf surface spectrum was 95.71%, and the spectral characteristics of the two after dimension reduction showed obvious separability. The first three principal components composite map could highlight the difference between leaf surface and leaf vein, and highlight the characteristics of leaf vein. The GAN-SA-UNet segmentation algorithm can capture the vein features in complex leaf texture images. The segmentation accuracy and intersection over union were 98.93% and 66.23%, respectively. Compared with the original model, they were increased by 0.18 percentage points and 4.21 percentage points, respectively. The inference time of single image was 4ms. It showed strong generalization ability and efficient and accurate recognition ability in the verification test of different producing areas, parts, grades and types of tobacco leaves.