This project aims to develop a robust model for multi-class classification of blood cell images while optimizing accuracy and computational efficiency.
In this project, we addressed the challenge of classifying 96x96 RGB blood cell images. Our goal was to create a model capable of generalizing well on unseen and potentially augmented datasets, all while keeping an eye on computational efficiency and storage requirements. The entire project has been tested on the CodaBench platform, while the notebooks have been run on Google Colab.
| Model | Augmentation | Validation Accuracy | Test Accuracy |
|---|---|---|---|
| Custom VGG16 | No | 98.41% | 98.58% |
| Custom VGG16 | Yes | 22.41% | 22.83% |
| MobileNetV3 | No | 92.64% | 91.05% |
| MobileNetV3 | Yes | 74.16% | 74.41% |
| ResNet50V2 | No | 88.71% | 90.13% |
| ResNet50V2 | Yes | 41.56% | 43.65% |
| InceptionV3 | No | 18.14% | 17.47% |
| InceptionV3 | Yes | 14.30% | 14.72% |
| EfficientNetV2M | No | 97.83% | 97.66% |
| EfficientNetV2M | Yes | 98.66% | 98.41% |
A thorough and extensive explaination of all the steps and phases of our project can be found in the report.
- Andrea Giangrande
- Marta Giliberto
- Emanuele Greco