An automatic wheat ear counting model based on the minimum area intersection ratio algorithm and transfer learning

Abstract

An accurate prediction of the number of wheat ears per unit area is crucial to wheat production, current research on the number of wheat ears per unit area is still dominated by manual surveys. In order to develop a rapid and low-cost algorithm for counting wheat ears on wider genotypes, reduce the influence of occlusion and adhesion on the recognition results of wheat ears image, this research proposes an automatic smartphone-based wheat ear counting method. By adopting the minimum area intersection ratio (MAIR) feature extraction algorithm, the transfer learning technology is used to achieve automatic wheat ear counting based on the YOLOv5 model. The K-means feature extraction algorithm is optimized, and the MAIR feature detection algorithm is adopted to improve recognition accuracy, achieving an optimized wheat ear density recognition of R2 = 0.87, compared with the model without the MAIR algorithm, the recognition accuracy was significantly improved. In the efficiency and performance test on a wider range of genotypes, the accuracy and efficiency of the model based on YOLOv5s training are optimal, the optimal depth transfer model achieves R2 = 0.95, the results demonstrated that transfer learning could significantly improve the recognition accuracy of the recognition model. In addition, the accurate density identification was achieved in the whole interval from the filling period to maturity. Therefore, the wheat ear number prediction based on MAIR algorithm and transfer learning technology can improve the accuracy of wheat ear counting.