Development of weed and crop detection algorithms provides theoretical support for weed control and becomes an effective tool for the site-specific weed management. For weed and crop object detection tasks in the field, there is often a large difference between the number of weed and crop, resulting in an unbalanced distribution of samples and further posing difficulties for the detection task. In addition, most developed models tend to miss the small weed objects, leading to unsatisfied detection results. To overcome these issues, we proposed a pixel-level synthesization data augmentation method and a TIA-YOLOv5 network for weed and crop detection in the complex field environment.
The pixel-level synthesization data augmentation method generated synthetic images by pasting weed pixels into original images. In the TIA-YOLOv5, a transformer encoder block was added to the backbone to improve the sensitivity of the model to weeds, a channel feature fusion with involution (CFFI) strategy was proposed for channel feature fusion while reducing information loss, and adaptive spatial feature fusion (ASFF) was introduced for feature fusion of different scales in the prediction head.
Test results with a publicly available sugarbeet dataset showed that the proposed TIA-YOLOv5 network yielded an F1-scoreweed, APweed and mAP@0.5 of 70.0%, 80.8% and 90.0%, respectively, which was 11.8%, 11.3% and 5.9% higher than the baseline YOLOv5 model. And the detection speed reached 20.8 FPS.
In this paper, a fast and accurate workflow including a pixel-level synthesization data augmentation method and a TIA-YOLOv5 network was proposed for real-time weed and crop detection in the field. The proposed method improved the detection accuracy and speed, providing very promising detection results.