ZHANG Hetao;ZHAO Chunjiang;WANG Chuanyu;GUO Xinyu;LI Dazhuang;GOU Wenbo;College of Information Engineering, Northwest A&F University;Beijing Research Center for Information Technology in Agriculture;Beijing Key Labof Digital Plant,National Engineering Research Center for Information Technology in Agriculture;

Wheat ears;;Convolutional neural network;;Feature extraction;;Feature fusion;;Loss function;;Wheat ears recognition and detection model

In order to improve the recognition and detection accuracy of convolutional neural network for wheat ears, an improved recognition and detection model YOLOv5-αTB was proposed based on the YOLOv5 detection model. Transformer module was added to the end part of feature extraction network to enhance the extraction of color, texture and feature of wheat ear image. In the feature fusion part, the Path Aggregation Network(PANet) was replaced by the weighted Bidirectional Feature Pyramid Network(BiFPN) to further optimize the multi-scale feature fusion. Aiming at the limitation of IoU, α-CIOU was introduced to enhance the effect of boundary box regression. The precision, recall, and average precision(AP) obtained by YOLOv5-αTB detection model on the test set are 99.95%, 81.86%, and 88.64%, respectively. Compared with the traditional YOLOv5 model, the average precision of the model was improved by 2.92 percentage points. Compared with the results of manual counting, the detection precision of the model was about 97.00%.