This project is part of the exam of Artificial Neural Networks and Deep Learning course at Politecnico di Milano, and it corresponds to the first challenge. The aim of the project is to classify species of plants, which are divided into categories according to the species of the plant to which they belong. Being a classification problem, given an image, the goal is to predict the correct class label.
The typical deep learning pipeline will be used, building a CNN from scratch and implementing transfer learning and fine tuning with 6 CNN: ResNet152V2, InceptionV3 InceptionResNetV2, VGG16, DenseNet201, Xception, EfficientNetB0. Tensorflow and in particular Keras framework will be used.
It is possible to check all the details of the project from the pdf in this repository, together with the notebooks of the models.
The dataset consists of 3543 images of plants distributed in 8 different species.
In order to solve the unbalanceness of the classes, the initial dataset has been oversampled. An alternative of this approach could have been undersampling, but for the current dataset this option is not efficient because the dataset contains few images.
The splitting of the initial dataset has been performed obtaining 2 sets: a training set with 80% of the initial data and a validation set with 20% of the initial data. The test set was not created and splitted because the model will be tested on a hidden and private test set.
Data augmentation is a technique used to increase the diversity of the training set by applying random (but realistic) transformations. It has been the first technique used in order to increase the number of images in the training set. Many transformations have been used: Height Shift Range, Width Shift Range, Zoom Range, Horizontal Flip, Shear Range, Fill Mode.
The models proposed are 7: a network from scratch, ResNet152V2, InceptionV3, InceptionResNetV2, VGG16, DenseNet201, Xception, EfficientNetB0. All the model except the network from scratch have been used as feature extractor, using the transfer learning paradigm.
Implementing a network from scratch helped to explore the problem and understand what are the most suitable structures for a network based on this specific problem of classification. The performances where not low, since the accuracy on the validation reached 78%. Transfer learning was the tool that obtained the best result, since very powerful model has been used to extract in an optimized and efficient way the features form the images, being already trained with a hige corpus of images. All the models that ahve been used as feature extractor plus a fully connected part related to the problem performed quite well, having accuracy above 85%. EfficientNet was the best among the other, with high performances, around 91% in the accuracy of the validation set, and 82% on the hidden test set.
Starting from scratch, it was hard to make a good network, due to the limited number of images. Therefore, using transfer learning was a good choice in terms of training time, as well as initialization, since those networks already had optimized weights. However, very powerful networks are not always the way to go because they really depend on the problem faced.
| Cognome | Nome | Codice Persona | |
|---|---|---|---|
| La Greca | Michele Carlo | michelecarlo.lagreca@mail.polimi.it | 10864460 |
| Giuffrida | Andrea | andrea.giuffrida@mail.polimi.it | 10643540 |
| Nicolis | Nicholas | nicholas.nicolis@mail.polimi.it | 10867841 |