Mushrooms have long been considered a valuable food source, some species also have potential therapeutic benefits, while others contain dangerous toxins. In this context, machine learning and deep learning-based image recognition methods are becoming promising approaches to mushroom classification. Although deep learning models have shown high efficiency in this field, they still have some limitations, such as the risk of overfitting due to complex architecture, a large number of parameters, and significant computational costs. To overcome these challenges, this study proposes using the EfficientNetV2-S model combined with Grad-CAM techniques and a suitable hyperparameter optimization strategy to improve classification efficiency. Experimental results on publicly available mushroom image datasets show that the proposed model achieves superior performance compared to many other modern methods for the same problem. Specifically, the model achieved an overall accuracy of 0.98, while the precision, recall, and average macro F1-score all reached 0.98. These results confirm the effectiveness and reliability of the proposed model and demonstrate that combining model architecture tuning, feature visualization techniques, and appropriate optimization strategies not only helps to limit overfitting and reduce computational costs but also significantly enhances the practical applicability of the fungal classification system.
Kim et al. (Fri,) studied this question.