Early-stage detection of maize seed germination based on RGB image and machine vision

Abstract

Maize (corn) seed germination rate is an essential piece of information to reflect seed quality and its marketability. The widely used seed germination test is manual, inefficient, time-consuming (required 7 days), and error-prone. This study utilizes machine vision combined with characterization of sand deformation and crack formation during germination for early and automatic germination detection. Collected color (RGB) images of germination trays planted with maize seeds sown in preset patterns were preprocessed as regions of interest (RoI) for each seed for analysis. For each RoI, direct early germination prediction methods, namely, stripe band, boundary, and color were developed using different image processing operations. A total of 36 images (4 trays for 9 consecutive days) were used to test the three direct methods and their different combinations. Experimental results showed that the performance of stripe band + boundary + color combination was superior to each direct method, and the average precision, recall, and F1 value of germination detection were 73.5 %, 87.5 %, and 79.2 %, respectively. It was also found that the seed germination rate detected on the 4th day (92.4 %) of the germination test could determine whether it met the sowing requirements, significantly shortening (by 3 days) the standard germination procedure time. This study demonstrates that the stripe band + boundary + color method can be used as an efficient approach for automated germination rate detection of maize and other crop seeds.