A novel YOLOv8n-C3 algorithm for silkworm body lesion recognition in diverse environments

Aiming at the problems of low detection confidence of YOLOv8n model in diversified environments and insufficient detection ability of fine lesion regions, an improved YOLOv8-C3 algorithm is proposed in this paper. The algorithm improves the model's accuracy in recognizing and detecting lesion regions in silkworms by improving the C2f module in the YOLOv8n backbone network and introducing the CGLU module and CaFormer module, respectively. Specifically, the CGLU module optimizes the C2f module with finer channel attention, which significantly improves the model's ability in spatial feature extraction, enabling it to more accurately discriminate and emphasize key features in the image. The CaFormer module, on the other hand, combines the advantages of deep separable convolution and ordinary self-attention mechanisms to effectively reduce computational complexity and improve long-range feature capture, thus enhancing the model's ability to express multi-dimensional features and overall performance. In addition, the combination of these techniques significantly improves the generalization ability of the model, enabling the improved YOLOv8n algorithm to adapt to diverse detection environments. These diverse environments include: low light, dense scenes, image exposure, added noise data, occluded targets, and complex backgrounds, etc., which further improves the model's detection effectiveness and robustness in practical applications. In order to verify the effectiveness of the improved algorithm, this paper gradually adds CGLU and CaFormer modules to YOOv8n backbone network through ablation experiments to compare the performance of the original YOOv8n algorithm and the improved algorithm. The experimental results show that the average detection accuracy of the model (YOLOv8n-C2F-CaFormer) after introducing the CaFormer module (mAP@0.5). It increased from 86.2% to 87.9%, and precision and recall also increased to 87.7% and 79% respectively. After the introduction of CGLU module, the model (YOLOv8n-C2F-CGLU) reaches mAP@0.5 87.1%, precision and recall 88.2% and 78.5% respectively. When CaFormer and CGLU modules are introduced at the same time (YOLOv8-C3), the model mAP@0.5 with the highest value of 90.8% was achieved on the, and precision and recall reached 88.9% and 83.1% respectively. This result proves that the improved YOOv8n algorithm has significant advantages in improving the accuracy and robustness of silkworm disease detection. This work was supported in part by the Special Project of Guangxi Collaborative Innovation Center of Modern Sericulture and Silk under Grant 2023GXCSSC01, and in part by Guangxi Natural Science Foundation Joint Funding Project under Grant 2020GXNSFAA159172,and IoT and Big Data-based Branching Technology for Mulberry and Silkworm Cultivation Production Monitoring Platform 2022GCC010,and Re-search Basic Ability Improvement Project for Young and Middle-aged Teachers of Guangxi Universities 2024KY0627,2023KY0633 ,and Hechi University University-level Research Project2023XJYB010, 2024XJPT006 ,2024XJYB011 and 2024XJPT005 , and CollegeStudent Innovation and Entrepreneurship Project 202410605023.