基于纹理特征与深度学习的小麦图像中的穗粒分割与计数

Segmentation and counting of wheat spikes and grains based on texture features and deep learning

穗粒数是小麦产量构成的重要因素和估测产量的参数之一,传统的人工计数方法耗时费力,人为因素影响大。为了实现对小麦穗粒数的智能、快速监测,以百农307、新麦26、稷麦336这3个小麦品种为试验材料,利用智能手机于小麦灌浆后期拍摄麦穗图像,随后对麦穗图像进行预处理并归一化为480×480像素大小,结合深度学习和迁移学习机制,构建基于冻结-解冻机制的HRNet模型的小麦小穗图像分割计数深度学习模型,利用图像处理算法、小麦小穗图像纹理特征确定小穗像素数与穗粒数之间的关系阈值,构建小穗粒数预测模型,实现对小麦穗粒的预测计数。结果表明,对比同样采用冻结-解冻机制的PSPNet模型、DeeplabV3+分割模型、U-Net模型及无冻结解冻机制的HRNet模型,采用基于冻结-解冻机制的HRNet模型对小麦小穗的分割效果更优,且具有更好的鲁棒性,分割精确度为0.9594,平均交并比(mIoU)为0.9119,类别平均像素准确率(mPA)为0.9419,召回率为0.9419;通过3个不同品种小麦的麦穗图像对小穗进行计数,所得决定系数(R2)为0.92,平均绝对误差为0.73,平均相对误差为2.89%;籽粒计数的R2为0.92,平均绝对误差为0.43,平均相对误差为5.51%。由研究结果可知,基于冻结-解冻机制得出的HRNet模型的小麦小穗图像分割算法能够有效分割小麦图像中的小穗,并获得更加丰富的语义信息,可用于解决小目标图像分割困难及训练欠拟合问题,通过粒数预测模型可以快速、精确地对小麦的籽粒数进行预测,从而为小麦高效、智能化估产提供算法支撑。

The number of grains per ear is an important factor in the composition of wheat yield and one of the parameters for estimating wheat yield.The traditional manual counting method is time-consuming and labor-intensive,and human factors have a great influence.In order to realize the intelligent and rapid monitoring of the number of grains per ear,three varieties of Bainong 307,Xinmai 26 and Jimai 336 were used as test materials,and the wheat ear images were taken with a smart phone at the late stage of wheat grain filling.Based on the image processing technology,the wheat ear images were preprocessed and normalized to 480×480 pixels.Combining deep learning and transfer learning mechanisms,a HRNet wheat spikelet segmentation and counting deep learning model based on the freeze-thaw mechanism was constructed.Image processing algorithms and wheat spikelet texture features were used to determine the threshold of the relationship between the number of spikelet pixels and the number of grains per spike.The spikelet-grain number prediction model was constructed to realize the prediction and counting of wheat spikes.The results showed that compared with PSPNet,DeeplabV3+segmentation model,U-Net which also used freeze-thaw mechanism and HRNet without freeze-thaw mechanism,the HRNet model based on the freeze-thaw mechanism had a better segmentation effect on wheat spikelets,and had better robustness.The segmentation accuracy was 0.9594,the mean intersection over union(mIoU)was 0.9119,the mean pixel accuracy(mPA)was 0.9419,and the recall rate was 0.9419.The spikelets were counted by the images of three different wheat varieties.The determination coefficient(R2)was 0.92,the average absolute error was 0.73,and the average relative error was 2.89%.The R2 of grain counting was 0.92,the average absolute error was 0.43,and the average relative error was 5.51%.It shows that the HRNet wheat spikelet image segmentation algorithm through the freeze-thaw mechanism can effectively segment wheat spikelets and obtain richer semantic information,which can be used to solve the problems of difficult segmentation of small target images and training underfitting.The model can quickly and accurately predict the number of wheat grains,so as to provide algorithm support for efficient and intelligent yield estimation of wheat.