A Study on Deep Learning Based Classification of Flower Images

Abstract

Deep learning techniques are becoming more and more common in computer vision applications in different fields, such as object recognition, classification, and segmentation. In the study, a classification application was made for flower species detection using the deep learning method of different datasets. The pre-learning MobileNet, DenseNet, Inception, and ResNet models, which are the basis of deep learning, are discussed separately. In experimental studies, models were trained with flower classes with five (flower dataset) and seventeen (Oxford 17) types of flowers and their performances were compared. Performance tests, it is aimed to measure the success of different model optimizers in each data set. For the Oxford-17 data set in experimental studies; With Adam optimizer 93.14% in MobileNetV2 model, 95.59% with SGD optimizer, 92.85% with Adam optimizer in ResNet152v2 model, 88.96% with SGD optimizer, 91.55% with Adam optimizer in InceptionV3 model, 91.55% with SGD optimizer Validation accuracy of 87.66, InceptionResnetV2 model was 86.36% with Adam optimizer, 83.76% with SGD optimizer, 94.16% with Adam optimizer in DenseNet169 model and 90.91% with SGD optimizer. For the dataset named Flower dataset; With Adam optimizer 91.62% in MobileNetV2 model, 80.80% with SGD optimizer, 92.94% with Adam optimizer in ResNet152v2 model, 85.03% with SGD optimizer, 90.71% with Adam optimizer in InceptionV3 model, 82% with SGD optimizer, 62, InceptionResnetV2 model, 88.62% with Adam optimizer, 81.84% with SGD optimizer, 90.03% with Adam optimizer in DenseNet169 model, 82.89% with SGD optimizer. When the results are compared, it is seen that the performance rate of deep learning methods varies in some models depending on the number of classes in the data set, and in most models depending on the optimizer type.