Accurate Identification of Grade of Grape Damage by <i>Brevipalpus</i> spp. Based on the Improved YOLOv8n Model

Abstract

Brevipalpus spp. are widespread pests on wine grapes in northwest China and have generated a major threat to the local wine grape industry in recent years. We advanced the YOLOv8n model (object detection algorithm), termed SEM-YOLOv8n, to predict the degree of damage from these mites, and thereby provided the appropriate time for pest management. The damage symptoms of Brevipalpus spp. were classified into the following five grades: non-infested, slight, moderate, severe, and extremely severe; the pictures of different grades were structured into a self-constructed dataset. Regarding algorithm improvements, to improve the ability to recognize subtle differences between the various grades of damage symptoms in complex natural backgrounds, the EMA attention mechanism was introduced after the SPPF layer of the backbone network. Secondly, to address the problem of target omission caused by grapevine fruit overlapping, the MPDIoU loss function was used instead of the CIoU loss function. Finally, the Slim-Neck structure was adopted in the neck of YOLOv8n to generate a lightweight model. The experimental results showed that the improved model increased the mean accuracy by 1.1% and decreased the number of parameters by about 13.3% compared with the original model. Compared with the related authoritative YOLO series algorithms, the improved model proposed in this study had a better detection performance in terms of both the accuracy and model size.

Keywords

• <i>Brevipalpus</i> spp.
• degree of damage
• YOLOv8n
• attention mechanism
• loss function