基于改进Mask R-CNN的水稻茎秆截面参数检测方法

针对人工测量、统计作物茎秆显微切片图像中维管束数目、面积等关键参数主观性强、费时费力、效率低的问题，提出一种基于图像处理的水稻茎秆截面参数自动检测方法。首先构建了一个基于改进Mask R-CNN网络的水稻茎秆切片图像分割模型。网络以MobilenetV2和残差特征增强及自适应空间融合的特征金字塔网络为特征提取网络，同时引入PointRend增强模块，并将网络回归损失函数优化为IoU函数，最优模型的F1值为91.21%,平均精确率为94.37%,召回率为88.25%,平均交并比为90.80%,单幅图像平均检测耗时0.50 s,实现了水稻茎秆切片图像中大、小维管束区域的定位、检测和分割；通过边缘检测、形态学处理及轮廓提取，实现茎秆截面轮廓的分割提取。本文方法可实现对水稻茎秆截面面积、截面直径，大、小维管束面积，大、小维管束数量等6个参数的自动检测，检测平均相对误差不超过4.6%,可用于水稻茎秆微观结构的高通量观测。

Automatic Detection of Rice Stem Section Parameters Based on Improved Mask R-CNN

Addressing difficulties in manual measurement and statistics of key parameters like the number and area of vascular bundles in crop stem microsection images such as high subjectivity, large time, labor investment, and low efficiency, an automatic detection method of rice stem cross-section parameters based on image processing was proposed. First of all, an image segmentation model of rice stem slices based on the improved Mask R-CNN was built. The network adopted MobilenetV2 and residual feature enhancement and the adaptive space fusion feature pyramid network as the feature extraction network. In the meantime, the PointRend enhancement module was introduced, and the regression loss function of the network was optimized to IoU function. The F1 value of the optimal model was 91.21%; the average precision rate was 94.37%; the recall rate was 88.25%; the mean intersection over union was 90.80%; and the average detection time of a single image was 0.50s. It achieved localization, detection and segmentation of large and small vascular bundle areas in rice stem slice images. Through edge detection, morphological processing and contour extraction, the stem section contours were segmented and extracted. The method proposed herein realized automatic detection of six parameters, namely rice stem section area, section diameter, large and small vascular bundle area, and the number of large and small vascular bundles. The average relative error of detection was no higher than 4.6%. The method can also be used for high-throughput observation of rice stem microstructure.