Real-Time Wheat Unsound Kernel Classification Detection Based on Improved YOLOv5

China is one of the largest wheat production coun-tries in the world. The wheat quality determines the price and many other aspects. The detection meth-ods of wheat quality mainly depend on manual la-bor. It costs high amount of manpower and time, and the classification results are partly affected by different individuals. With the development of ma-chine vision, an automatic classification system was presented in this study. A wheat unsound kernel iden-tification method based on the improved YOLOv5 al-gorithm was designed by adding efficient channel at-tention (ECA). Compared with convolutional block at-tention module (CBAM) and squeeze-and-excitation network (SENet), the improved YOLOv5 algorithm was selected to fit the model better. The recognition results showed that YOLOv5 with the addition of the attention mechanism had a significant improvement in average accuracy over that without it. The most signif-icant improvement was observed with the addition of ECA-YOLOv5, with an average accuracy of 96.24%, a 10% improvement over the other two models, and a 13% improvement over the original YOLOv5. This satisfied the application requirements for detection of wheat unsound kernel.