Klasifikasi Spesies Jamur Menggunakan Metode Hibrida DenseNet-121 dan Support Vector Machine

Abstract

The public’s inability to distinguish between the visual characteristics of poisonous and edible mushrooms contributes to a high risk of wild mushroom poisoning. This study developed a mushroom image classification system using a hybrid method combining a Convolutional Neural Network (CNN) based on the DenseNet-121 architecture and a Support Vector Machine (SVM), which was tested through four comparative modeling schemes. The four schemes include: (1) training DenseNet-121 from scratch, (2) integrating DenseNet-121 with SVM without Transfer Learning, (3) Transfer Learning of DenseNet-121 using ImageNet weights, and (4) Transfer Learning + Fine-Tuning + SVM. Each scheme was evaluated using three batch size variations (4, 8, 16) and two optimizers (Adam and RMSProp). The results show that Scheme 4, using the Adam optimizer and a batch size of 4, achieved the highest average accuracy of 92.37%, far surpassing the baseline model in Scheme 1, which reached only 82.40%. The best model was then converted to the TensorFlow Lite format and implemented in an Android application. Testing on 180 new images (unseen data) recorded an accuracy of 94.44% with a total inference time of only 0.50 seconds per image. This study proves that the DenseNet-121 + SVM hybrid architecture with a fine-tuning strategy excels in mushroom classification and is suitable as a mobile-based educational tool for preventing poisoning.