基于可变形卷积的稻粒在穗计数方法

Method of Counting Rice Grains in Ears Based on Deformable Convolution

摘要:

水稻穗粒数快速获取对筛选高产、优质品种具有重要意义，针对脱粒计数破坏稻穗拓扑结构，无法用于其他表型参数测量等问题，提出一种稻粒在穗计数方法。将稻粒在穗计数视为密度预测问题，基于可变形卷积，设计稻穗图像特征提取骨干网络，用少量选取的范本稻粒和稻穗图像的特征相关性，通过特征相关层生成特征相关图，在特征相关图基础上，重用并级联图像特征，预测稻粒密度分布，进而通过密度图求和，获取计数结果。试验结果表明，本文方法具有较高的计数精度，测试样本稻粒计数平均绝对误差（Mean absolute error, MAE）、均方根误差（Root mean squared error, RMSE）和平均相对误差（Mean relative error, MRE）分别为4.71、6.92和2.9%，MRE仅比人工走查高0.7个百分点，与现有基准方法（FamNet、CSRNet和ICACount）相比，MRE分别降低9.9、8.6、11.6个百分点；用可变形卷积设计的稻穗图像特征提取网络能有效提高稻粒计数精度，在参数量接近的前提下，基于该网络的模型MAE和RMSE比ResNet-50分别低19.3%和12.9%，模型具有良好的拟合能力，决定系数R2达0.940 5；相同网络架构下，可变形卷积比常规卷积在稻粒计数MAE和RMSE上分别降低28.9%和22.0%，MRE下降1.6个百分点；图像特征重用对提高稻粒计数精度具有重要作用，使模型在测试集上的MAE和RMSE下降27.6%和22.1%，MRE下降2.2个百分点。该方法单幅稻穗图像处理时间为0.92 s，有效提高了工作效率，可为稻穗表型检测和平台设计提供技术参考。

Abstract:

Rapid acquisition of grain number in rice spike is important for screening high-yielding and high-quality varieties. Aiming to address the problems that threshing counting destroyed the topology of the rice spike and cannot be used for the measurement of other phenotypic parameters, a method for counting rice grains in the spike was proposed. Considering the in-situ counting of rice grains as a density prediction problem, based on deformable convolution, a backbone network for feature extraction of rice spike images was designed. With a small number of selected paradigms for feature correlation of rice grains and spike images, feature correlation maps were generated through feature correlation layers, and based on the feature correlation maps, the image features were reused and cascaded to predict the distribution of density of the rice grains, which was then summed up to obtain the counting results through the density maps. The test results showed that the method had high counting accuracy. The mean absolute error (MAE), root mean square error (RMSE), and mean relative error (MRE) of rice grain counts of the test samples were 4.71, 6.92, and 2.9%. respectively, with MRE being only 0.7 percentage points higher than that of the manual walk-through, and MRE reduction of 9.9, 8.6 and 11.6 percentage points compared with that of existing benchmark methods FamNet, CSRNet and ICACount. Rice spike image feature extraction network designed with deformable convolution can effectively improve the accuracy of rice grain counting. With a close number of parameters, the model-based on this network was 19.3% and 12.9% lower than that of ResNet-50 in MAE and RMSE, and the model had a good fit with coefficient of determination R2 of 0.940 5. Deformable convolution reduced 28.9% and 22.0% of rice grain count MAE and RMSE, and 1.6 percentage points of MRE than conventional convolution for the same network architecture. Image feature reuse played an important role in improving the accuracy of rice grain counting, and this treatment decreased the MAE and RMSE of the model on the test set by 27.6% and 22.1%, and the MRE by 2.2 percentage points. The processing time of single rice spike image of this method was 0.92 s, which effectively improved the work efficiency, and the research can provide technical reference for rice spike phenotype detection and platform design.