This project is an image classification project using a deep-learning network and using a transfer learning approach with MobileNetV2 architecture provided by Keras.
The main approach during this project was to use gradual transfer learning where for each training of the 3 phases we defined, we fine-tuned the model by incrementally unfreezing a certain number of layers. The project has 3 versions with small variations in the network architecture with both having a test accuracy >95%. Tha last version has class weights added to the model in order to deal with class imbalance within the dataset. It has a slightly better accuracy >96%.
The dataset comprise images of 450 bird species. 70,626 training images, 22500 test images(5 images per species) and 2250 validation images(5 images per species. This is a very high quality dataset where there is only one bird in each image and the bird typically takes up at least 50% of the pixels in the image. As a result even a moderately complex model will achieve training and test accuracies in the mid 90% range.
- You can find a link to the dataset used in ths project here .
- You can find a link to the code output, including history logs and model weights here .
- Kaggle version
- Colab version