README.md

---
license: agpl-3.0
datasets:
- openfoodfacts/front_image_classification
base_model:
- Ultralytics/YOLO11
metrics:
- accuracy
---

# Front image classification model

This model classifies Open Food Facts images into two classes:

- `front` (ID 0)
- `other` (ID 1)

Front images are the "default" image of a product, displayed on Open Food Facts product page. A front image is most of the time a photo of the front side of the product packaging.
It's useful to be able to detect front images so that we can update the front image with a newer version (when the packaging changes for example).

## Model Details

### Model Description

- **Developed by:** Raphaël Bournhonesque
- **Model type:** Image Classification
- **License:** AGPL 3.0
- **Finetuned from model [optional]:** Yolo11n-cls

## Uses

This model is intended to be used on Open Food Facts images only (images of food packaged products).

## Training Details

### Training Data

v1.0 of the [front_image_classification](https://huggingface.co/datasets/openfoodfacts/front_image_classification) dataset was used to train the model.

### Training Procedure

- Epochs: 100
- Image size: 448
- Albumentation augmentation

[This script](https://github.com/openfoodfacts/openfoodfacts-ai/blob/dbbec40a3d964124cd7c8d838023be4a10d6c0be/front-image-classification/train.py) was used for training the model.

The augmentation pipeline used for prediction:

```python
A.Compose(
        [
            A.LongestMaxSize(max_size=max_size, p=1.0),
            A.PadIfNeeded(min_height=max_size, min_width=max_size, p=1.0),
            A.Normalize(mean=DEFAULT_MEAN, std=DEFAULT_STD, p=1.0),
            ToTensorV2(p=1.0),
        ]
    )
```

For optimal performance, it is advised to keep the same preprocessing pipeline during inference.

## Evaluation

accuracy: 0.9525

## Export

An ONNX export can be found in `weights/model.onnx`.