Recognition and Diagnosis of Mulberry Diseases Based on YOLOv8-GSBN

Highlights The BoTNeT backbone network is invoked in YOLOv8 to combine convolutional operations with self-attention mechanisms to augment the model’s ability to detect localized features such as diseases. The neck structure of YOLOv8 is improved, and a lightweight GSConv-Slim-Neck network is introduced to reduce the weight of the model while taking into account the accuracy of mulberry disease recognition. Experiments on a self-made mulberry disease dataset demonstrate that YOLOv8-GSBN achieves an improvement in both recognition accuracy and model size compared to other algorithms. Abstract. Manual identification of mulberry diseases is inefficient and often inaccurate, creating challenges for prevention and control and ultimately causing economic losses for farmers. Therefore, we propose a deep learning-based method for mulberry tree disease recognition and prevention, comprising a user-intelligent mobile front-end, a cloud server back-end platform, a disease recognition model running on the platform, and a disease database. Detecting various mulberry tree diseases is challenging because of their similar visual features and subtle differences against the background. Therefore, we propose an improved Convolutional Neural Network model based on YOLOv8s: YOLOv8-GSBN as the disease recognition model for this method. We introduce the BoTNet structure into the original YOLOv8s to augment the model’s feature-capture range and global feature extraction, and integrate GSConv-Slim-Neck by replacing convolutional modules with GSConv and the C2f module with the cross-stage VoV-GSCSP, thereby improving recognition accuracy while maintaining a lightweight structure. To verify the model’s effectiveness, we also constructed a dataset spanning six categories, including healthy mulberry and healthy mulberry leaves, and obtained a total of 2,295 high-quality mulberry disease images for model training through data balancing and other data augmentation techniques. In the comparison experiments with other algorithms, the YOLOv8-GSBN model proposed in this study achieves a precision of 90.1%, a recall of 79.7%, and an mAP of 88.5% on the mulberry pest dataset. Compared with the original YOLOv8, it improves precision, recall, and mAP by 4.4%, 3.7%, and 5.4%, respectively, while reducing FLOPs and parameters by 11.4% and 3.6%. In addition, the model achieves an inference speed of 147 FPS, fully satisfying the real-time detection requirements . These results demonstrate that the proposed YOLOv8-GSBN model not only enhances recognition accuracy but also effectively reduces computational complexity, making it more suitable for deployment on mobile and edge devices. In conclusion, the mulberry disease identification and control method proposed in this article can efficiently and accurately identify disease species and provide mulberry farmers with control methods to augment the intelligence level of mulberry disease control. Keywords: Deep learning, Disease detection, Mulberry tree disease, YOLOv8s.