AI-YOLACT Model for Automatic Severity Grading of Microbial-Based Anthracnose Infection in Camellia Leaves

Anthracnose, a notable ailment, has a substantial impact on the development of Camellia oleifera. It is imperative to accurately assess the severity of this illness in order to apply pesticides precisely and implement scientific management effectively. This study introduces a classification model called AI-YOLACT, which utilizes artificial intelligence to assess the severity of microbial-based anthracnose infection on Camellia leaves. The AI-YOLACT model is meant to automatically and quickly perform this assessment. The Swin Transformer is employed in the central network of AI-YOLACT to extract features. The self-attention mechanism in the transformer design significantly improves the model’s ability to extract features by utilizing a global receptive field and shifted window characteristics. An innovative approach called a weighted bidirectional feature pyramid network is proposed to effectively combine feature information from multiple scales. This technique enhances the model’s capability to accurately identify targets of varied sizes, hence enhancing detection accuracy. The activation function of the original model, ReLU, is replaced by the more potent nonlinear HardSwish function. Due to the fact that HardSwish is not entirely disconnected in the negative area, it demonstrates increased resistance to noise in the input data. Due to the intricate background and foreground details in natural environment photographs, the resilience of HardSwish aids the model in effectively managing such scenarios, hence enhancing accuracy. The technique of transfer learning was utilized to perform experiments on a dataset aimed at assessing the extent of microbial-induced anthracnose infection in Camellia leaves. The findings of the ablation experiment demonstrate that the AI-YOLACT model achieves a mAP_75 of 86.8%, which is a 5.7% enhancement compared to the baseline. Additionally, it achieves a mAP_all of 78.3%, showing a 2.5% improvement, and a mAR of 91.6%, indicating a 7.9% enhancement. Comparative trials have shown that AI-YOLACT surpasses SOLO (Segmenting Objects by Locations) in terms of both accuracy and speed. Additionally, AI-YOLACT’s detection speed is twice as quick as the Mask R-CNN technique. AI-YOLACT’s performance was validated in 36 outdoor grading experiments, resulting in a 94.4% accuracy in assessing the degree of anthracnose on Camellia leaves. The average absolute K-value error was 1.09%. The Camellia-AI-YOLACT model, introduced in this study, demonstrates exceptional precision in assessing the quality of Camellia leaves and the degree of anthracnose lesions. Consequently, it enables the automated grading of anthracnose severity. It offers technical assistance for accurate disease management in Camellia plants, hence enhancing the automation and intelligence of anthracnose diagnosis.