Identification of varieties of wheat seeds based on multispectral imaging combined with improved YOLOv5

Abstract The identification of seed variety is important in wheat production because the growth and yield are highly related with its variety. Traditional discrimination methods for wheat seed varieties were suffered with time consuming and contamination. In this study, multispectral imaging technology combined with improved YOLOv5s was proposed. Three optimal spectral bands images were selected from all 19 bands using Genetic algorithm and confusion matrix and then the new images were fused with the three band images. The improved YOLOv5s with CBAM module was developed to train the identification model of wheat varieties using fusion images. The experimental results showed that the average precision of the model from proposed method in test set reached 99.38% that were better than the traditional YOLOv5s model. Meanwhile, the evaluation indexes of the model such as P/%, R/%, F1/% and mAP/% were all higher than 90%. The results showed that the method could be used for wheat variety identification rapidly and non-destructively.