Image-based recognition of edible weeds: creating a new dataset, classification studies with different deep learning models

In this study, classification analyses were conducted to recognize edible wild herbs that grow naturally in the wild. Various classification algorithms were employed, and their accuracy rates were compared. Images of ten different edible herb species were processed using image processing techniques. Due to the limited number of samples in the dataset, data augmentation techniques were applied, and the images were expanded using the ImageDataGenerator. The augmented images were then classified using pre-trained deep learning models, namely DenseNet121, MobileNetV2, ResNet50, and VGG16. The models were evaluated using both statically and dynamically augmented training datasets, and their accuracy scores and training times were compared. According to the results, the highest validation accuracy was achieved by the ResNet50 model with 96.96%, while the DenseNet121 model yielded the best test accuracy at 96.97%. The findings demonstrate that pre-trained deep learning models provide effective results with small datasets. The overall aim of the study is to contribute to the documentation of these plants and to support their increased recognition in future research efforts.