Leaf Classification:

An application of deep reinforcement learning

This model trains on grayscale images of 99 different species of leaves. Approximately 1580+ images in all and 16 images per species. For full description of the dataset see kaggle.

Requirements:

• python 3.5
• tensorflow
• keras
• theano
• install remaining package dependencies via pip install -r requirements.txt

How to Train:

• Clone this repo
• Download the dataset
• Navigate to repo directory and run python preprocess.py -d 'your_path_to_data'
  • This takes all leaf images stored in your_path_to_data and processes them to be 32x32 grayscale images.
  • Processed image files now located in the repo directory under processed
• Next navigate to whichever model you wish to train and run python learn.py -m Train

Model Descriptions:

• Deep Recurrent Reinforcement Network
  • Located in /reinforcement
  • The model simulates a game in which the play has 99 possible moves/actions.
  • Given an image of a leaf, the player must make one move. If the move matches with the leaf's species ID, then a positive reward is given. If not, a negative reward is given. This emulates a sort of "flash card" study game in which the learner looks at the image, makes a decision, and during training immediately discovers if the decision is accurate or not.
  • Model includes Long Short Term Memory (LSTM) components.
  • During training a target network and training network are used as a form of competitive learning.
• Deep Convolutional Neural Network (with Images)
  • Located in /cnn
  • Inputs: leaf image
  • Processes through two convolutional layers followed by two connected layers
  • Incorporates batch normalization, dropout regularization, and SGD
• Deep Convolutional Neural Network (with Feature Vectors)
  • Located in /1d-nn
  • Inputs: leaf features (margins, shapes, textures) formatted as 1-dimensional vector
  • Processes through two convolutional layers followed by two connected layers
  • Incorporates batch normalization, dropout regularization, and SGD
• Deep Highway Network
  • Located in /highway-net
  • Inputs: leaf image
  • More information soon