zhaoyu19920930/kaggle_flowers_classification_keras

Kaggle_Flowers_Classification_Keras

Descrition

The dataset is from Kaggle's Flowers Recognition. The goal is to classify five kinds of flowers (chamomile, tulip, rose, sunflower, dandelion) by raw image.

Dataset

The dataset contains 4242 images of flowers. The pictures are divided into five classes: chamomile, tulip, rose, sunflower, dandelion. For each class there are about 800 photos. Photos are not high resolution, about 320x240 pixels.

Preprocessing

1. Resize all the input images to 224x224.
2. 0.8 training samples && 0.2 validation samples

Model

There are three kinds of network architectures I used for this dataset.

1. The first model is build from scratch which has four layers.
2. The second model is build by pre-trained model VGG19 (freezing first 5 layers && include_top=False) and customed fully connected layer.
3. The third model is build by pre-trained model ResNet-50 ((freezing the first layer && include_top=False) and customed fully connected layer.

Result

Model	Accuracy for validation samples
Built from scratch	0.72
Built by VGG19	0.4
Built by ResNet-50	0.92

Notes

• Computing: Google Colab Tesla K80 GPU
• Python version: 3.6.3
• Using packages
  1. Keras (tensorflow.python.keras) for building models
  2. OpenCV (cv2) for processing images
  3. sikit-learn (sklearn) for train_test_split