Study on Wheat Spike Automatic Detection Method Based on Improved YOLOv8n

ZANG Hecang; ZHOU Meng; WANG Yahui; PENG Yilong; ZHAO Qing; ZHANG Jie; LI Guoqiang

doi:10.15933/j.cnki.1004-3268.2025.07.017

Abstract：In wheat breeding,spike number is the key index to evaluate wheat yield.Timely and accurate detection of wheat spike number has important practical significance for early prediction of yield.In actual production,the method of artificial field investigation and statistics of wheat spikes is time-consuming and laborious.Therefore,this paper proposed an automatic wheat spike detection method based on improved YOLOv8n.Firstly,HGNetV2 was used to improve the network structure to enhance the expression ability of small target wheat spike feature;Secondly,deep separable convolution and pointwise convolution were introduced to improve the computational efficiency and counting performance of the model;Finally,the loss function was improved to optimize the model,accurate determination of wheat ear position and category information was achieved.The test results showed that the accuracy of the improved YOLOv8n in wheat spike detection task was 93.7%,which was 6.5 percentage points higher than that of YOLOv8n.Compared with YOLOv5s and YOLOv8x,the improved YOLOv8n increased by 9.7 percentage point and 0.5 percentage point,which could detect wheat spike images in field complex situations,and had better computer vision processing effect and performance evaluation detection effect.This method can accurately detect the number of small target wheat spikes,and better solve the problem of occlusion and overlapping of wheat spikes.