Optimasi Algoritma SVM dengan Ekstraksi Fitur Warna pada Klasifikasi Biji Kopi Sangrai

Abstract

Manual classification of roasted coffee beans is often hampered by subjectivity and inconsistency. This study aims to compare the performance of the Support Vector Machine (SVM) algorithm in classifying the ripeness of roasted coffee beans using two different feature approaches: raw pixel features and extracted hybrid features. The research focuses on finding the best hyperparameter combination for each approach and determining which method yields superior classification performance. The dataset used consists of 900 augmented coffee bean images, evenly distributed across three classes (Dark, Light, Medium). In the first approach, an SVM model was trained directly on flattened raw pixel data. In the second approach, an SVM model was trained using combined features extracted via a Convolutional Neural Network (CNN) and Color Histogram. The experimental results show a significant performance difference. The SVM model using raw pixel features achieved a maximum accuracy of 88.33% with the best parameters {kernel: 'rbf', C: 10, gamma: 0.01}. Meanwhile, the hybrid model utilizing feature extraction from the deeper_wider CNN architecture and color histograms drastically improved performance, reaching an accuracy of 98.33% with parameters {kernel: 'rbf', C: 10, gamma: 0.1}. These results demonstrate that employing high-level feature extraction through CNN is significantly superior to using raw pixels for the task of roasted coffee bean classification.