基于小麦群体图像的田间麦穗计数及产量预测方法

在田间小麦测产时，需人工获取田间单位面积内的麦穗数和穗粒数，耗时耗力。为了快速测量小麦田间单位面积内的产量，该文利用特定装置以田间麦穗倾斜的方式获取田间麦穗群体图像，通过转换图像颜色空间RGB→HSI，提取饱和度S分量图像，然后把饱和度S分量图像转换成二值图像，再经细窄部位粘连去除算法进行初步分割，再由边界和区域的特征参数判断出粘连的麦穗图像，并利用基于凹点检测匹配连线的方法实现粘连麦穗的分割，进而识别出图像中的麦穗数量；通过计算图像中每个麦穗的面积像素点数并由预测公式得到每个麦穗的籽粒数，进而计算出每幅图像上所有麦穗的预测籽粒数，然后计算出0.25 m2区域内对应的4幅图像上的预测籽粒数；同时根据籽粒千粒质量数据，计算得到该区域内的产量信息。该文在识别3个品种田间麦穗单幅图像中麦穗数量的平均识别精度为91.63%，籽粒数的平均预测精度为90.73%；对3个品种0.25 m2区域的小麦麦穗数量、总籽粒数及产量预测的平均精度为93.83%、93.43%、93.49%。运用该文方法可以实现小麦田间单位面积内的产量信息自动测量。

Method for wheatear counting and yield predicting based on image of wheatear population in field

Abstract: At present, the wheatear number and grain number for unit area wheat in field can be measured when predicting yield. Generally, phenotype parameters should be obtained by manual count technique. It is time-consuming and needs great effort. In order to quickly measure the yield of unit area wheat in field, the wheatear population image was obtained by tilting the wheatear with specific device in field. The contour information on wheatear image in field was collected. Firstly, the color space of wheatear population image was converted from RGB (red green blue) to HSI (hue saturation intensity) for the sake of improving the uniformity of image color. Then the saturation component of image was extracted from the HSI color space of wheatear population image. Binary image of the saturation component of image was obtained by using image binary algorithm, morphological opening operation and removing of small regions algorithm. Then binary image was smoothed by linear mean filtering algorithm. The adherent and narrow part was removed by setting distance threshold between boundary points. Then the adhesive wheatear in image was judged out by its boundary and region characteristic parameters from the binary image. The boundary characteristic parameters included the length of entire boundary and the angle of on boundary point. The region characteristic parameters included the region area and shape factor of region and duty cycle of convex closure. Then image edge of adhesive wheatear was smoothed by using concave domain smoothing method. Then concave points on the boundary of adhesive wheatear were extracted by using included angle method and area method. The concave point pairs were found by 6 matching principles of concave points. The adhesive wheatear in image was segmented by connecting concave point which was already detected and matched on the binary image boundary. The separated wheatears and non-adhesive wheatears were superimposed on a binary image. The connected regions on the binary image were marked by image labeling algorithm. The number of wheatears in one binary image was counted. And the total number of wheatears in 0.25 m2 area was obtained by summing the number of wheatears in corresponding 4 wheatear images. Meanwhile, the area pixels number of each wheatear in binary image was extracted. The grain number prediction formula of wheatear in image was obtained by the linear relationship between actual grain number and area pixels number of pre-marked wheatear. Then the grain number of each wheatear in binary image was forecasted by using grain number prediction formula. The total grain number of wheatear in one image was obtained by summing the grain numbers of each wheatear in binary image. The total grain number of wheatear in 0.25 m2 area was obtained by summing the grain numbers in corresponding four wheatear images. The 1 000-grain weight of 3 varieties of wheat which included Suke wheat 1, Yang wheat 22 and Su wheat 188 was measured respectively. Finally the yield of wheat in 0.25 m2 area was calculated according to the 1 000-grain weight and the total grain number of wheatear. Compared with the actual wheatear number, grain number in a wheatear image and actual yield information of wheat in 0.25 m2 area, the experiment results manifest that the average identification precision of the wheatear number in a wheatear image for 3 varieties of wheat is 91.63%, and the average prediction precision of the grain number in a wheatear image for 3 varieties of wheat is 90.73%. And the average prediction precision of the total wheatear number, total grain number and yield of wheat in 0.25 m2 area for 3 varieties of wheat are 93.83%, 93.43% and 93.49%, respectively. The automatically predicting yield information of wheat in unit area can be realized by using wheatear image features method.