This repo contains the code for conference paper titled Flower Species Recognition System using Convolutional Neural Networks and Transfer Learning, by I.Gogul and V.Sathiesh Kumar, Proceedings of ICSCN-2017 conference, IEEE Xplore Digital Library.

• Pretrained state-of-the-art neural networks are used on University of Oxford's FLOWERS17 and FLOWERS102 dataset.
• Models used - Xception, Inception-v3, OverFeat, ResNet50, VGG16, VGG19.
• Weights used - ImageNet
• Classifier used - Logistic Regression
• Tutorial for this work is available at - Using Pre-trained Deep Learning models for your own dataset

Update (16/12/2017): Included two new deep neural net models namely InceptionResNetv2 and MobileNet.

• Theano or TensorFlow sudo pip install theano or sudo pip install tensorflow
• Keras sudo pip install keras
• NumPy sudo pip install numpy
• matplotlib sudo pip install matplotlib and you also need to do this sudo apt-get install python-dev
• seaborn sudo pip install seaborn
• h5py sudo pip install h5py
• scikit-learn sudo pip install scikit-learn

• This project used Windows 10 for development purposes and Odroid-XU4 for testing purposes.

MIT License

• Organize dataset - python organize_flowers17.py
• Feature extraction using CNN - python extract_features.py
• Train model using Logistic Regression - python train.py

The below tables shows the accuracies obtained for every Deep Neural Net model used to extract features from FLOWERS17 dataset using different parameter settings.

• Result-1

  • test_size : 0.10
  • classifier : Logistic Regression

• Result-2

  • test_size : 0.30
  • classifier : Logistic Regression