Objective YOLOv5l algorithm model was introduced and improved to realize accurate, efficient and nondestructive detection of rice panicles in field environment.

Method Taking rice in the field as the research object, rice image samples were collected by digital single-mirror reflex camera. The original image data were augmented and expanded offline after manual labeling, so as to construct an image data set for field rice. The YOLOv5l algorithm was improved adaptively, the effective channel attention (ECA) mechanism was put in front of the spatial pyramid pooling (SPP) layer and in the cross-stage-partial-connections (CSP) layer, and a comparative experiment was conducted. The optimal algorithm was selected as the benchmark model to carry out attention mechanism and data-enhanced ablation experiments, and the optimal performance model was obtained by testing. The improved YOLOv5l was compared with YOLOv5l, YOLOv5x, SSD and Faster R-CNN.

Result In the improved rice recognition framework of YOLOv5l, placing ECA before the network SPP layer resulted in better performance. Using test set images to verify the model, the average accuracy of recognition results was 93.63%, the average recall rate was 90.94%, and the overall average accuracy reached 95.05%. Compared with the non-fused YOLOv5l algorithm, the average accuracy of the improved YOLOv5l algorithm was 3.03 percent higher and the detection rate was 8.20 frames per ms faster. Compared with the YOLOv5x algorithm, the average precision of the improved YOLOv5l algorithm was improved by 0.62 percent, the detection rate was faster by 5.41 frames per ms, and the memory occupation was reduced by 74.1 MB. The results showed that the comprehensive performance of the improved YOLOv5l algorithm was better than other algorithms in rice panicle detection in the field.

Conclusion It is feasible to introduce the improved YOLOv5l algorithm into rice panicle detection in field environment. The algorithm has high accuracy, fast detection speed and small memory occupation, which can avoid the subjectivity of traditional manual detection and is of great significance for rice panicle detection and non-destructive yield estimation.