Convolutional Neural Network: Leaves Classification

Artificial Neural Networks and Deep Learning competition 2021/2022 - Politecnico di Milano.

Authors: Fabio Tresoldi and Mirko Usuelli

1. Introduction

The given dataset is unbalanced with respect to theclasses distribution, then we decided to apply a StratifiedSampling procedure in order to preserve the full dataset proportions both in training, validation and testing.This structural characteristic suggests the usage of theCategorical Crossentropy as loss function, whereas as metrics we kept into account accuracy and F1-score, being thelatter more meaningful for unbalanced categories.

2. Data Pipeline

2.1. Data Augmentation

Due to the reduced dimension of the proposed datasetcompared to the parameters complexity of the underlying architecture, Data Augmentation purpouse consists in solvingthis issue by generating at run time several new transformedimages which allow a better generalized performance of the model. Starting from Traditional Transformations (2.1.1), we tried to furtherly improved the quality of classification by introducing some Advanced Transformations (2.1.2).

2.1.1. Traditional Transformations.

Rotating, Zooming,Flipping, Brightness, Shifting.

2.1.2. AdvancedTransformations.

CutMix,MixUp,CutOut.

However, these last Advance Transformations did notincreased the overall performance in our specific case, butindeed, they made the training process slower and not asaccurate as without them, therefore we decided to discardthese techniques.

2.2. Pre-Processing

Since for the Features Exctractor (section 3.1) werelied on State-of-Art architectures through Fine Tuning, the data pre-processing function adopted is thesame suggested by the corresponding model contained in tensorflow.keras.applications.

3. Convolutional Neural Network

3.1. Features Extractor

We started comparing the most famous architectures showed during the course lectures through Transfer Learning and a simply GAP + Softmax layers as baseline classifier. Then we plotted both Categorical Crossentropy loss andAccuracy for the validation and training set (dashed line) within 10 epochs:

As shown in the above figure, the best models were EfficientNet and ResNet. While the latter converges faster, the former isable to reach a higher peak of performance in both loss and metrics as epochs go further. Thus, we have chosen EfficientNet as Features Extractor.

3.2. Classifier

Starting from a general classifier model, with the main layers shown during the course, we analyzed the performance of each of them through Transfer Learning (see the report for details). Here we show the final model we identify through several testing comparisons:

4. Performance

Finally, we generated the confusion matrix to identifythe correctness of classification for each class with F1-score, precision and recall, on the dataset split in 60% for trainingand 20% each for validation and testing:

• Accuracy: 99.77%
• Precision: 99.78%
• Recall: 99.75%
• F1-score: 99.76%

5. Leadboard Evaluation

• Development phase accuracy : 94.91%
• Final phase accuracy : 94.53%