A backlight and deep learning based method for calculating the number of seeds per silique

Abstract

Rapeseed is one of the most important oil crops in the world, and the rapeseed yield is increasing every year. The number of seeds per silique is one of the critical factors in the rapeseed yield. Studies that examine the number of seeds per silique have an important influence on the rapeseed yield measurement and the breeding of high-yield rapeseed varieties. Image-analysis-based seed counting methods have the advantages of being fast, accurate, and convenient. Based on the light-transmitting characteristic of siliques, this study used the backlight method to obtain images of the siliques' inner seeds. Three methods, the OTSU, Faster-RCNN, and DeepLabV3+, were used for the rapeseed segmentation and counting under different light intensities. The results showed that the silique images obtained under the 18,600 l× light intensity were the most conducive to seed segmentation. Under this condition, the DeepLabV3+ method had the best accuracy for the segmentation and counting of rapeseed. The Recall and F1-score were greater than 91% and 94%, respectively. • Provides non-destructive method for calculating the number of seeds per silique. • The number of seeds in silique can be achieved directly based on backlight shooting. • The deep learning model performs better in seed number calculation. • The calculate accuracy of model is greater than 93%.