Implementasi Hybrid Feature Fusion menggunakan ResNet50 dan RGB-HSV untuk Klasifikasi Indikasi Tingkat Kepedasan Cabai Berdasarkan Karakteristik Visual

Abstract

Chili is a horticultural commodity with diverse visual characteristics and different levels of pungency, requiring an objective and consistent classification method. A challenge in chili classification is the similarity of visual characteristics across categories and the separate use of color features or deep learning features in previous studies, resulting in less comprehensive information representation. This study aims to implement hybrid feature fusion using ResNet50 and RGB-HSV color features to classify indications of chili pungency into Hot, Medium, and Mild categories based on visual characteristics. The novelty of this study lies in integrating deep features from ResNet50 with RGB-HSV features into a single feature representation, as ResNet50 can capture complex visual patterns, while RGB-HSV explicitly provides color information. This study employed a quantitative approach with an experimental methodology based on image processing and deep learning. The dataset consisted of 1,592 chili images, including 416 Hot, 514 Medium, and 662 Mild images. The data were divided at a 70:15:15 ratio into 1,114 training, 239 validation, and 239 testing images. Preprocessing included background removal using rembg, resizing images to 224 × 224 pixels, and extracting mean and standard deviation features from RGB and HSV channels. The model achieved 96.23% accuracy, 96.10% precision, 96.08% recall, and a 96.08% F1-score, outperforming the original image model. These findings indicate that integrating deep and color features improves chili classification performance.