Using (deep) convolutional neural networks to classify a dataset of leaf photos. A 2019 final graduation project (optional exceptional learning achievement) in German language during high school.
We first introduce the topics of machine learning, artificial intelligence and artificial neural networks and explain the theoretical and mathematical foundations of computer vision and convolutional neural networks. Then, we focus on solving the problem of classifying images of leafs from a small labeled dataset in a controlled environment with state-of-the-art neural network architectures and test them against each other as well as novel self-generated architectures. The development is a success and lands near-perfect classification scores for unseen test data.
Please visit the final paper in GERMAN language for an extensive documentation of the project and research results.
Supervised by Hans Dietmar Jäger.
Studying image recognition using artificial (convolutional) neural networks, using TensorFlow and Keras frameworks for the implementations in Python.
The dataset is self-collected. Please ask for permission for further usage.
Working, but unmaintained since early 2019.
- Implementation: Python, TensorFlow, Keras.
- Architectures Tested: VGG, Nguyen, 5 filters to 64 filter, variation of kernel size, experimentation with dropout and dense layers as well as multi-block architectures.
- Number of parameters: 204,950 - 88,962,178.
- Number of layers: 3 - 16.
- Maximum accuracy: 100%.