Semi-automated, machine vision based maize kernel counting on the ear

Abstract

In synchrony with the overall trend toward automation in plant phenotyping, two semi-automatic machine vision methods were devised targeted at counting maize kernels directly on de-husked ears. The ears represented row morphologies ranging from straight to curved, they featured missing kernels, underdeveloped kernels, and broken kernels, but displayed no disease or mould. The first method mimicked a common manual field method of estimating the total ear kernel count, based on counting the number of rows and multiplication by the number of kernels per row. Similarly, in this paper, the operator manually counted the number of rows, and also manually counted the number of kernels in a row image within an (operator determined) quasi-cylindrical mid-section of the ear. The total ear kernel count was then estimated by multiplying the number of rows by the number of kernels per row, yielding full ear extrapolation by multiplication by the ratio between the total ear length and the length of the quasi-cylindrical mid-section. This full ear image based approach achieved a kernel counting error ranging from −7.67% (under-count) to +8.60% (multi-count) among 23 maize ears. The second method only observed a fixed quasi-cylindrical mid-section of the ear. Image frames were acquired of each individual row of kernels located in the quasi-cylindrical mid-section, yielding kernel maps. Among 12 maize ears, the kernel missing error ranged from 0 to 4.24% and the multi-count error ranged from 0 to 1.92%. In total, 41 existing kernels were missed and 25 kernels were multi-counted among 2713 kernels counted.