Abstract
Peppers are widely cultivated and consumed in Bangladesh, with varieties such as capsicum (red, yellow, green), local chili (red, green), Shimla, and Bombay chili. They are loaded with vitamins (C, A, B6, K, and E), antioxidants (carotenoids, flavonoids), and minerals (potassium). Accurate breed classification is vital for agriculture, quality control, and market organization. Although deep learning has been used in the classification of peppers in other studies, relatively few varieties and limited datasets were analyzed. This paper fills in this gap by presenting a new balanced dataset that consists of 1,000 images on 10 pepper types and quantitatively comparing the performance of three models, such as VGG16 (Simonyan & Zisserman, 2014), ResNet50 (He et al., 2016), and YOLOv10. Both VGG16 and ResNet50 are classical models for image classification, while YOLOv10 is a general-purpose object detection model which has been used here for the purpose of classifying pepper breeds. For instance, they reported that ResNet50 and VGG16 obtained 96% classification accuracy, and YOLOv10 achieved 94.25% accuracy with better real-time speed. Precision, recall, and F1-score analysis showed that YOLOv10 performed more stably in visually similar breeds (e.g., half-red vs. green chili), although with less overall precision. Statistical confidence intervals assured that these findings were satisfactory. The results reveal trade-offs in model accuracy, speed, and deployment feasibility, and show that self-localization-based and event recognition-based models play complementary roles in agricultural automation