Intel-images-classification

A comparison between Transfer Learning and custom Convolutionnal Network to classify images.

TS;WR (Too short; Will Read)

Read the full report (in french) here: Report.pdf

Situation

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.

Action

I built two models:

• Custom CNN: a simple CNN
• EfficientNet: a state-of-the-art model for image classification wiht ImageNet weights.

Results

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.

Custom CNN

EfficientNet