Rose Breed Detection Using Transfer Learning Method

Abstract

For flower cultivation, flower creation and flower business, breed identification of a particular flower and providing the description of that breed with planting recommendation and maintaining ways are very beneficial. There are so many local flowers are available in Bangladesh, rose is one of the most common and wanted flower. Not only in Bangladesh roses are the most acceptable flower all over the world. Roses are most preferable flowers for decoration; besides that, it has many more uses. Rose breed identification will have impressive effect on floriculture and flower business because roses have largest involvement in flower business. Though floriculture is improving day by day, but there is no available approach for breed identification of specific flower rather than various flower recognition approaches. In this project, I presented a model based on transfer learning techniques to identify breed of roses from pictures. For image processing and classification of flowers, resources are not sufficient, so there was great necessity of a broad dataset with huge number of pictures to train the model. I have collected 1939 real pictures of five different breeds of rose and I have created 9306 and 388 images for training dataset and testing dataset accordingly. I have used four transfer learning techniques in my project, which are Inception V3, ResNet50, Xception and VGG16. VGG16 scored the maximum accuracy of 99%, which is an outstanding performance. Among all flower related works, breed detection of a particular flower is the first approach based on my knowledge.