A Study on CNN Based Transfer Learning for Recognition of Flower Species

Abstract

The flower that is one of the plant organs, is essential element of the ecological order. Flowers have been used in many areas that are beneficial to humans. There exist about four hundred thousand varieties of flowers known today. It is a difficult task to distinguish flowers from each other due to their similarity in shape and color. Flower classification is a challenging problem due to the high variety of shapes, color distribution, lighting conditions and deformation of exposure. It becomes more difficult to distinguish flowers that are similar in color and shape to each other with the human eye for some images. It takes remarkable training for humans to correctly distinguish between particular species, and often very specific morphological features are the only thing that distinguishes closely related species. CNN models have been recently used by researchers in many classification problems to eliminate the need for manual features. In this study, CNN-based transfer learning methods are studied for recognition of flower species. Popular pretrained learning techniques which are VGG16, VGG19, SqueezeNet, DenseNet-121, DenseNet-201, and InceptionResNetV2 are conducted for classification of flower species. Their classification performances are compared on same flower dataset in experimental results. It was observed that the InceptionResNetV2 model gives superior results than other models in experiments. The highest accuracy (92.25%) is obtained with the InceptionResNetV2 model for flower dataset.