Root-YOLOv7: Multi-scale adaptive object detection and grading of root-knot nematode disease

Root-knot nematodes can infect over 2000 plants, which causes significant economic losses. Rapid and accurate detection of root-knot nematode disease is the key to screening resistant varieties and evaluating the effect of prevention and control. To address the challenge of detecting root-knot nematode disease caused by dense root knots, multiple fibrous roots, and small root knots, this paper takes cucumber as the research object to propose a detection and grading model Root-YOLOv7. Specifically, the backbone network of YOLOv7 is restructured, which enables the proposed model to better capture object information at different scales through a hierarchical structure and a self-attention module based on shifted windows. Additionally, combined with the Wise-IoU loss function, the proposed model can adaptively adjust the weight of the overlapping part of the object box, which solves the problem that bounding box regression cannot be optimized effectively when detecting low-quality objects. Furthermore, an improved head network structure is proposed to compute the attention weights by capturing the cross-dimension interaction of the root knot feature between the spatial and channel dimensions. To verify the effectiveness of the proposed model, the performance of Root-YOLOv7 is compared with typical object detection models. Experimental results show that the AP@0.5 of Root-YOLOv7 reaches 87.40%, which is 72.67%, 5.60%, 5.28%, 9.68%, 5.83%, and 7.45% higher than Faster R-CNN, RT-DETR, YOLOv5, YOLOv6, YOLOv7, and YOLOv8, respectively. The proposed approach is expected to reduce the workload of plant pathologists and provide technical support for the cultivation of plant varieties with disease resistance.