Chili leaf segmentation using meta-learning for improved model accuracy

Abstract

Recognizing chili plant varieties through chili leaf image samples automatically at low costs represents an intriguing area of study. While maintaining and protecting the quality of chili plants is a priority, classifying leaf images captured randomly requires considerable effort. The quality of the captured leaf images significantly impacts the development of the model. This study applies a meta-learning approach to chili leaf image data, creating a dataset and classifying leaf images captured using mobile devices with varying camera specifications. The images were organized into 14 experimental groups to assess accuracy. The approach included 2-way and 3-way classification tasks, with 3-shot, 5-shot, and 10-shot learning scenarios, to analyze the influence of various chili leaf image factors and optimize the classification and segmentation model's accuracy. The findings demonstrate that a minimum of 10 shots from the meta-test dataset is sufficient to achieve an accuracy of 84.87% using 2-way classification meta-learning combined with the mix-up augmentation technique.