Background: The number of kernels per ear is one of the major agronomic yield indicators for maize. Manual assessment of kernel traits can be time consuming and laborious. Moreover, manually acquired data can be influenced by subjective bias of the observer. Existing methods for counting of kernel number are often unstable and costly. Machine vision technology allows objective extraction of features from image sensor data, offering high-throughput and low-cost advantages.
Results: Here, we propose an automatic kernel recognition method which has been applied to count the kernel number based on digital colour photos of the maize ears. Images were acquired under both LED diffuse (indoors) and natural light (outdoor) conditions. Field trials were carried out at two sites in China using 8 maize varieties. This method comprises five steps: (1) a Gaussian Pyramid for image compression to improve the processing efficiency, (2) separating the maize fruit from the background by Mean Shift Filtering algorithm, (3) a Colour Deconvolution (CD) algorithm to enhance the kernel edges, (4) segmentation of kernel zones using a local adaptive threshold, (5) an improved Find-Local-Maxima to recognize the local grayscale peaks and determine the maize kernel number within the image. The results showed good agreement (> 93%) in terms of accuracy and precision between ground truth (manual counting) and the image-based counting.
Conclusions: The proposed algorithm has robust and superior performance in maize ear kernel counting under various illumination conditions. In addition, the approach is highly-efficient and low-cost. The performance of this method makes it applicable and satisfactory for real-world breeding programs.
Keywords: Adaptive threshold; Computer vision; Counting; Kernel recognition; Local Maxima.
© The Author(s) 2020.