Classification of fungi images using different convolutional neural networks

Abstract

Fungi offer vital solutions to humanity through roles in medicine, agriculture, and ecological balance while presenting potential threats. They have yielded antibiotics, food fermentation, and nutrient recycling however, fungal infections, crop diseases, and spoilage highlight their dark side. Therefore, it is important to identify fungi to harness their potential benefits and mitigate threats. Offering quick and accurate identification through image classification improves the aforementioned features. Therefore, this study classified images of five types of fungi using convolutional neural networks (CNN). Initially, dataset distribution was observed, and it was identified that there was a class imbalance in the dataset. To address this issue, data augmentation technique was used. Several preprocessing techniques were also applied to understand the model training behavior with their application. Then the images were rescaled into six different resolution combinations such as original images, low-resolution images, high-resolution images, a mix of original and low-resolution images, a mix of original and high-resolution images, and a mix of low and high-resolution images. Then these data were trained using 13 pre-trained CNN models such as Xception, VGG16, VGG19, InceptionResNetV2, ResNet152, EfficientNetB6, EfficientNetB7, ConvNeXtTiny, ConvNeXtSmall, ConvNeXtBase, ConvNeXtLarge, ConvNeXtXLarge, BigTransfer (BiT). To evaluate these models, accuracy, macro average precision, macro average recall, macro average f1- score, and loss learning curve assessment were used. According to the results, the BiT model preprocessed with normalization, which used a mix of original and high-resolution images, performed the best, producing a model accuracy of 87.32% with optimal precision, recall, and f1-score. The loss learning curve of the BiT model also depicted a low overfitting aspect proving the model’s optimal behavior. Therefore, it was concluded that the BiT model with the mix of original and high-resolution data can be used to detect fungi efficiently.