A comparison between Transfer Learning and custom Convolutionnal Network to classify images.
Read the full report (in french) here: Report.pdf
Transfer learning is used to solve image classification problems when we don't have time or computational power. But does it work all the time?
Here, I compare EfficientNet performances and a custom Convolutional Neural Network to classify images.
We have 5 classes:
- Mountain
- Sea
- Building
- Forest
- Iceberg
I used this dataset from kaggle: Intel-image.
I built two models:
- Custom CNN: a simple CNN
- EfficientNet: a state-of-the-art model for image classification wiht ImageNet weights.
EfficientNet performed poorly, with an accuracy of less than 20% while my custom CNN yielded a result of 83%.
Note that, the dataset is not good quality and some images are mislabeled, which I believe, affected my model's performances.