Abstract: To address the issues of strong subjectivity and low detection efficiency in traditional pepper (Capsicum annuum L.) disease grading methods, this study proposed a machine vision-based semantic segmentation model for automated rapid grading and identification of anthracnose-infected pepper fruits.Under controlled enclosed environments,sunlight was simulated, and images of healthy fruits and four disease severity levels across different pepper varieties were collected. Principal component analysis was employed to reduce redundant image features, extracting three key color components (Cb, Cr, R) with a cumulative contribution rate of 95%. Model 1 (Decision Tree), model 2 (Naive Bayes), model 3 (SVM), and model 4 (KNN) were trained. Model 1 (Decision Tree) demonstrated the shortest training time and highest precision, establishing it as the optimal prediction model for anthracnose disease grading. It required low computational resources and occupied minimal memory, facilitating future edge deployment. Model 1 achieved precision rates of 90.3%~98.2% for pepper fruits and 75.3%~80.7% for disease spots. Its recall rate for anthracnose disease grading was 73.3%~93.3%, with the recall rate for healthy peppers (Level 0) exceeding 90.0%. The prediction results of model 1 showed high consistency with manual annotations across all disease levels, verifying its applicability in automated disease monitoring systems as a replacement for manual visual grading methods.

Key words: pepper (Capsicum annuum L.) fruit, machine vision technology, anthracnose, disease grading