Algorith for acquiring multi-phenotype parameters of soybean seed based on OpenCV

Abstract: Phenotypic trait parameters of soybean seeds were greatly contributed to the soybean breeding. Deep Learning, particularly Convolutional Neural Networks (CNN), has been introduced into the acquisition and analysis of plant phenotypes in recent years. However, the existing deep learning algorithms cannot fully meet the high requirement of large-scale production, due to the less phenotypic traits and a high-cost CNN training. A convenient high-throughput approach is required to accurately obtain the phenotypic trait parameter of soybean seeds. In this research, an acquisition algorithm was proposed to extract the multiple-phenotypic trait parameters of soybean seeds using the OpenCV image processing library and computer vision. The image collection of soybean seeds was easily and rapidly completed using the mobile phone photography during the soybean seed test. All seeds of each soybean plant to be detected were also photographed as an image. Furthermore, the grayscale histogram of the original image was firstly established to automatically generate a binary graph. The morphological processing was then used to enhance the image details and remove the image noise. The improved watershed algorithm was used to extract the contours of soybean seeds in the image. The circularity was introduced to evaluate the seed contour. The secondary contour segmentation with a higher grayscale threshold was performed for the special seed adhesion in the small areas. The seed circularity was also introduced to identify the incomplete soybean seeds, according to the contour information. The proportion of abnormal RGB areas was calculated to determine the sick soybean seeds with the epidermis discoloration. The lengths of long and short axis, cross-section area, and circumference of soybean seed were calculated using the ellipse fitting and scale bar conversion. The CSV table files were used to store for all the phenotypic trait data of each soybean seed and the average phenotypic trait data of all soybean seeds in each image. The soybean plants were also sorted to optimize the soybean plant seeds with the excellent phenotype for the breeding experiment design. The acquisition algorithm was utilized to identify the soybean seeds, and then to extract the phenotype parameters of soybean seeds. The results show that the recognition rate of the total soybean seeds in each image reached 98.4%, the correct recognition rate of the damaged and diseased soybean seeds was 95.2%, as well as the calculation accuracies of the long and short axis length of soybean seed reached 96.8%, and 95.8%, respectively. The parallel computation of the algorithm was implemented to create the multiple processes. By introducing 8-process parallel calculation, the image processing time was reduced by two-thirds compared to single-process calculation. Thus, the proposed algorithm was easily parallelized to quickly realize the accurate acquisition of multiple phenotypic trait parameters of soybean seeds, including the circumference, area, long/short axis length, roundness, and RGB value. At the same time, an accurate identification was achieved in the good and damage soybean seeds, including the incomplete and the sick seeds.