Combining Convolutional Neural Networks for Fungi Classification

Deep learning approaches have shown exceptional efficacy in challenges related to the categorization of images. Nevertheless, the practical use of these methods in classifying fungi faces difficulties due to the distinctive features of fungal morphology and the scarcity of annotated training data. Hence, this study introduces an innovative and inclusive methodology that utilizes the spatial transformer network technique to analyze fungi thoroughly feature alterations. The fungal characteristics are then subjected to processing by integrating four networks. The combined convolutional neural networks are enhanced with adaptive layers, convolutional operations, kernel sizes, dropblock mechanisms, and residual blocks. These components collaborate harmoniously via the concatenate function in the feature mapping process. In that order, the experimental findings demonstrate notable training accuracies, namely 91.89%, 98.24%, 98.49%, and 98.92%. Furthermore, the classification of fungi showcases remarkable precision, achieving high accuracies of 98.91%, 82.50%, 94.11%, 100.0%, and 87.43% for Absidia, Aspergillus, Fusarium, Penicillium, and Rhizopus, respectively. Additionally, the recall performance stood at 100.0%, 87.61%, 96.78%, 100.0%, and 88.45% for Absidia, Aspergillus, Fusarium, Penicillium, and Rhizopus, respectively. The findings of this study suggest that the suggested deep learning approach has considerable promise in developing a reliable system for identifying fungus species.