<![CDATA[Accurate classification of Capsicum annuum varieties is crucial for food industry applications and agricultural research. Traditional manual classification methods are time-consuming, subjective, lack detail, and are prone to human error, requiring computer vision to automate them. This study presents learning in the form of automatic classification of nine diverse Capsicum annuum varieties using transfer learning with the MobileNetV3 architecture, which is designed to achieve high accuracy and be computationally energy efficient. The dataset consists of 4,500 images (training, testing, and validation) of 9 chili varieties: bell pepper, curly chili, cherry pepper, chiltepin, Hungarian wax, jalapeno, marconi, pequin, and Thai chili. This dataset goes through quality control, one of which is dataset balancing. The model in this study has also been optimized with Adam (Adaptive Moment Estimation). Model interpretation is also improved through Grad-CAM visualization, and model robustness has also been validated using cross-validation 5 times. This model achieved performance with a training accuracy of 97.2%, a testing accuracy of 95.1%, and a validation test of 94.8%, where 5-fold cross-validation showed consistent results (94.23% ± 1.45%). Grad-CAM analysis showed that this model focuses on structural features such as shape, surface texture, and color patterns. With the successful development of an AI system that can automatically identify chili varieties with an accuracy of 95.1%. This system works well in real conditions (90.6% accuracy) and is practical for use in agriculture and food processing. This technology can help farmers and food companies or lay people to sort chilies automatically, reduce costs, and improve quality control.]]>
Copyrights © 2025
