A Mobile App for the Identification of Flowers Using Deep Learning

Flowers are admired and used by people all around the world for their fragrance, religious significance, and medicinal capabilities. The accurate taxonomy of these flower species is critical for biodiversity conservation and research. Non-experts typically need to spend a lot of time examining botanical guides in order to accurately identify a flower, which can be challenging and time-consuming. In this study, an innovative mobile application named FloralCam has been developed for the identification of flower species that are commonly found in Mauritius. Our dataset, named FlowerNet, was collected using a smartphone in a natural environment setting and consists of 11660 images, with 110 images for each of the 106 flower species. Seventy percent of the data was used for training, twenty percent for validation and the remaining ten percent for testing. Using the approach of transfer learning, pre-trained convolutional neural networks (CNNs) such as the InceptionV3, MobileNetV2 and ResNet50V2 were fine tuned on the custom dataset created. The best performance was achieved with the fine tuned MobileNetV2 model with accuracy 99.74% and prediction time 0.09 seconds. The best model was then converted to TensorFlow Lite format and integrated in a mobile application which was built using Flutter. Furthermore, the models were also tested on the benchmark Oxford 102 dataset and MobileNetV2 obtained the highest classification accuracy of 95.90%. The mobile application, the dataset and the deep learning models developed can be used to support future research in the field of flower recognition.