小麦麦穗几何表型测量的精准分割方法研究

Study on precise segmentation method for geometric phenotype measurement of wheat ear

[目的]小麦麦穗表型获取涉及麦穗到籽粒不同几何尺度的参数精确测量,本文针对麦穗籽粒图像分割粘连现象,研究达到像素级别的精准分割算法,并基于该方法给出籽粒的基本几何参数。[方法]田间随机采集小麦麦穗,对采集的麦穗标本获取表型信息并采集图像,进行数据增广和标注,构建1个包括深度残差网络(deep residual network,ResNet)、区域建议网络(region proposal networks,RPN)和全卷积网络(fully convolutional networks,FCN)的实例分割算法MaskR-CNN,对训练集图片进行迭代训练获得模型。[结果]测试集测量结果表明,在测试麦粒上获得的籽粒像素测量平均精度(average-precision,AP)值为0.85,F1(F1-measure)值为0.830,对麦穗长度测量穗长的平均绝对误差为3.30mm,平均相对误差为3.40%,宽度测量的平均绝对误差为0.72mm,平均相对误差为4.10%,综合测量误差为3.75%,试验结果显著优于最大类间方差法(OTSU)以及全卷积网络。通过对特征提取网络层数的修改在处理速度上达到4.26FPS(frames per second),对比FCN处理速度提升了8.5倍。[结论]利用MaskR-CNN分割方法得到1个对整株麦穗和单个籽粒进行目标定位、目标检测和实例分割为一体的端到端、像素级的分割模型,可以对麦穗及部分籽粒进行精确的几何表型测量。

[Objectives] The acquisition of wheat phenotype involves an accurate measurement of wheat ear and grain parameters at different geometric scales.It can segment the adhesion phenomenon of wheat grain image and study the precise segmentation algorithm to achieve pixel level.The basic geometric parameters of the kernel are given based on this method.[Methods] Wheat ears were randomly collected in the field.We obtained phenotypic information on wheat ears and collected images.Then we performed data augmentation and labeling to build an instance segmentation algorithm Mask R-CNN including deep residual network(ResNet),region proposal networks(RPN),and fully convolutional networks(FCN).We iteratively trained the training set image to obtain the model.[Results] The prediction results of the test set showed that the average accuracy of the grain pixel obtained on the test grain(average-precision,AP)was 0.85,the F1(F1-Measure)value was 0.83,and the average absolute error of the wheat ear length was 3.30 mm,the average relative error was 3.40%,the average absolute error of the width was 0.72 mm,the average relative error was 4.10%,and the comprehensive measurement error was 3.75%.The test results are significantly better than the maximum inter-class variance method(OTSU)and the full convolution network.By modifying the number of feature extraction network layers,the processing speed reached 4.26 FPS(frames per second),and the processing speed of the FCN was improved by 8.5 times.[Conclusions] A segmentation model was obtained by using the Mask R-CNN segmentation method,which can target the whole wheat and individual kernels,target detection,and instance segmentation.This is an end-to-end,and pixel-level segmentation model that accurately measures the geometric phenotype of wheat and some grains.