An improved YOLOv5-based algorithm for small wheat spikes detection

Wheat spike detection is of great research importance for wheat yield estimation as well as wheat quality management. For the detection of wheat spikes, object detection methods based on machine learning are often employed. The accuracy of wheat spikes photographs is significantly hampered by the images’ small size, extreme density, and heavy overlapping. For the features of wheat spikes images, we present an enhanced YOLOv5-based small wheat spikes detection technique in this research. This approach addresses the issue of erroneous and missed detection in the wheat spikes detection process. Specifically, by introducing the application of SPD-Conv in the wheat spikes detection algorithm, the performance of small-size object detection is improved and the missed detection of wheat spikes is reduced. By introducing the CA attention mechanism in the neck of the small wheat sheaf detection algorithm, thus improving the accuracy of object feature information extraction and detection, the problem of wrong and missed detection of small-sized wheat sheaves is effectively improved. By introducing an efficient RepGFPN module to replace the C3 module in the wheat spikes detection algorithm, the feature extraction capability of the model for small-sized wheat spikes is improved, and better detection accuracy is achieved. The experimental outcomes demonstrate that the enhanced algorithm for detecting wheat spikes can increase detection precision of wheat spikes with an mean average accuracy (mAP) of 94.3%\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\%$$\end{document}, which is better than the general object detection model.