<![CDATA[General Background Accurate identification of edible and poisonous mushrooms is critical for food safety because high visual similarity among species often causes misclassification. Specific Background Deep learning with transfer learning using Convolutional Neural Networks has been widely applied for image-based mushroom classification, particularly through pretrained architectures such as VGG-16 and ResNet50. Knowledge Gap Nevertheless, limited comparative evidence exists regarding which architecture provides more stable and balanced performance when applied to relatively small and diverse mushroom image datasets. Aims This study compares VGG-16 and ResNet50 transfer learning models for binary mushroom toxicity classification using the Kaggle Edible and Poisonous Mushroom Images dataset consisting of 2,820 images from 47 species. Results Using a 70:15:15 training, validation, and testing split with standardized preprocessing and data augmentation, the fine-tuned VGG-16 model achieved 96% test accuracy with a loss of 0.1671, while the ResNet50 model reached 92% accuracy with a loss of 0.2991. Both models obtained a ROC AUC value of 1.000, although VGG-16 demonstrated more balanced precision, recall, and F1-scores across classes. Novelty This research presents a direct and systematic comparison of two widely used pretrained CNN architectures under identical experimental settings. Implications The findings support automated mushroom toxicity identification to assist safer mushroom consumption decisions. Highlights: The strongest model achieved 96% accuracy with lower classification loss. The alternative model produced lower accuracy under identical conditions. Both approaches reached perfect ROC AUC with differing class balance. Keywords: Deep Learning, Convolutional Neural Network (CNN), VGG-16, ReNet50, Mushroom Image Classification]]>
Copyrights © 2026
