Optimized DenseNet Architectures for Precise Classification of Edible and Poisonous Mushrooms

Abstract

Abstract Background The subtle differences between edible and toxic mushroom species make classification difficult. Traditional methods often result in errors which led to misclassifications and conventional machine learning models often struggle in feature extraction due to subtle differences in mushroom species. Deep learning models, such as DenseNet architectures, offer potential solutions, but due to model complexity, deep architecture and large number of parameters these models suffer from overfitting and computational costs. These can be handled by optimizing the model. This study’s primary goal is to enhance the precision and reliability of mushroom classification through deep learning by enhancing the DenseNet-121 structure. This is done through the implementation of more sophisticated model regularization techniques and automating the hyperparameter optimization process. The study intends to show how model architectural changes and optimization approaches can offer solutions to issues like overfitting and significant resource expenditure on computation and ultimately improve mushroom classification systems for efficiency and efficacy. Methods The study analyzes the basic DenseNet-121 model as well as a modified DenseNet-121 with frozen upper layers which preserve important lower level features. Automated hyperparameter tuning is done with KerasTuner, while dropout and weight decay regularization methods are used to control overfitting. Evaluation metrics include accuracy, precision, recall, F1-score, confusion matrices, and other graphical methods. Results The enhanced DenseNet-121 model surpasses the standard DenseNet-121 in mushroom classification. While DenseNet-121 obtained an overall accuracy of 0.90 along with a macro average precision, recall, and F1-score of 0.90 $$-$$ - 0.91, Modified DenseNet-121 achieved 0.97 for all those metrics. The improvements increased the precision and recall for all the classes which means the model has less trustability and more accuracy in classification. Conclusion The study demonstrates the effectiveness of architectural modifications and regularization strategies in improving model performance. Despite problems such as possible over-reliance on pre-trained features and computational complexity, the modified DenseNet-121 is useful for accurate mushroom classification. Future study could look into improving freezing procedures and lowering computational demands to extend applicability.

Keywords

• DenseNet-121
• Modified DenseNet-121
• Mushroom classification
• Deep learning
• Machine learning
• Image classification