This project was developed as part of an Artificial Neural Networks and Deep Learning course at Politecnico di Milano (2021/2022). It was organized as an Codalab competition among students groups.
The goal was to perform a classifiction of leaves based on their images. The dataset was composed of 17728 leaf images of 13 different species of plants. In other words, given an image, the goal was to predict the correct species of the plant (class label).
The zipped dataset can be found here in file dataset.zip. Additionally, also the subset of data used in the TTA tests (TTA_tests.ipyinb) is made available in the TTA-dataset.zip.
To run the final_model.ipynb and TTA_tests.ipynb notebooks, download the datasets from the link in the Dataset section and unzip them to the directory of your choice. Then, adjust the paths to the datasets specified in the code notebooks.
The final model submitted by the team can be found in final_model.ipynb file. The model uses transfer learning with pre-trained VGG16 extended with one fully connected layer. Since we did not expect the VGG16 to match our classification problem, we fine-tuned the VGG16’s convolutional layers starting from the 14th layer.
The details of the model are described in the report.pdf available in this repository. To see how we addressed problems such as overfitting, class imbalance, and how we tuned the hyperparameters, please refer to the report.