Research On Early Germinated Wheat Detection Based On Deep Learning

Wheat is the third largest grain in the world after rice and corn,and China is the largest wheat producer.As a major grain crop in China,wheat plays a vital role in the development of the national economy and society.Early germinated wheat detection is crucial for the statistics of wheat germination rate and breeding,but traditional early germinated wheat quantity statistics are based on manual statistical methods.In practice,the small size and large number of wheat seeds lead to problems such as huge workload,time-consuming and susceptibility to personal subjective factors.In addition,there is no public dataset for early germinated wheat detection at present.In response to the problem of low efficiency of the manual counting method of total wheat seeds during early germination,this paper proposes a deep learning-based wheat germination detection method,and the one-stage algorithm model YOLOv5 s is selected as the basic research model.In order to further improve the detection accuracy of the model,the YOLOv5 s model was improved,and the recognition accuracy of the improved model for wheat seeds during early germination was improved by2%.In addition,this article also uses image processing technology to process wheat images and videos,generating corresponding visual images of the growth of wheat seeds in a certain time and space,and developing a graphical interface for wheat germination detection.The main work of this article is as follows:(1)The Wheat Germination Detection(WGD)was constructed.According to the requirements of the model and the detection target,the collected wheat germination images and video data are pre-processed,image enhancement and expansion,and then the processed wheat image is labeled with label Img,a data annotation tool,to construct a WGD dataset for wheat germination detection.(2)Three strategies are proposed to improve the wheat germination detection algorithm based on YOLOv5 s.In view of the above background,this paper has made three improvements to YOLOv5s: First,integrate the Swin Transformer structure into the tail end of the model’s trunk and neck networks,improve the trunk and neck networks,and enhance the information content of the feature map.This strategy improves the m AP of the model by0.6 percentage points;The second is to improve the feature fusion method of the model neck network based on the idea of Bi FPN network,strengthen the transmission of feature information on the network,and increase m AP by 0.4 percentage points;The third is to introduce an attention mechanism to further improve the neck network,improving its feature fusion efficiency and anti-interference ability,and improving the detection accuracy of germination seeds by 0.2 percentage points.(3)Based on the improved YOLOv5 s wheat germination detection method.Based on the above research,this paper integrates three improvement strategies,and ultimately improves the accuracy of the YOLOv5 s model by 1.6 percentage points,with a maximum m AP of 94.9%.The improved YOLOv5 s model was applied to the rapid detection of wheat seeds during early germination,and the results were displayed through a graphical interface,achieving automatic detection of wheat germination.