Using feature extraction with deep learning in flower classification

Deep learning techniques are widely used for image recognition and classification problems. Deep learning architectures have evolved to include more layers and become more robust models for classification problems. In this study, four deep learning models were fine-tuned to classify flowers into 20 categories. Features extracted by deep networks randomly trained from 20 different flower categories were classified using support vector machines. The dataset used was originally composed of plant families found on campus. Deep learning methods were implemented in Matlab. The dataset contained 20,000 images in total, 1,000 from each class. The model was trained using randomly selected images to prevent bias. The remaining images were used in the testing phase. Elements were selected from the dataset according to the training rate parameter. When the training rate exceeds 10%, the classification accuracies of ShuffleNet, SqueezeNet, and ResNet-50 models exceed 95%. However, GoogleNet's classification accuracy exceeds 95% after a 25% training rate. Confusion matrix metrics, one of the comparison metrics used in classification, were also evaluated according to the increase in training rate for four models.