Automatic count of wheat ears in field wheat by improved YOLOv7

Abstract

Considering the difficulty of counting wheat sheaves in the field, this paper proposes an improved Yolov7 (YOU ONLY LOOKCE version 7) model for the automatic counting of wheat sheaves in the field. Based on Yolov7, the method adds a simple parameter-free attention module (SimAM) and full-dimensional dynamic convolution (ODConv), which can enhance the dimensional interactivity of the backbone network in extracting features. By introducing a centralised feature pyramid (CFP) into the neck structure, a comprehensive and differentiated feature representation can be effectively obtained. The improved Yolov7 model improves the applicability of automatic wheat counting and allows for better suppression of useless information in complex field environments. Several models were selected for comparative testing in the collected wheat head dataset, and the results showed that the improved Yolov7 achieved an average accuracy of 96.5%, outperforming other target detection models and allowing more accurate identification of wheat spike counts.