---
library_name: transformers
tags:
- image-classification
- vision-transformer
- vit
---

# Model Card for Fruit Classification ViT Model

## Model Details

### Model Description

This model is a **Vision Transformer (ViT)** fine-tuned for fruit image classification. It classifies images into six categories:

- Banana  
- Mango  
- Orange  
- Pitaya  
- Pomegranate  
- Tomatoes  

The model is based on transfer learning using a pretrained ViT architecture and has been fine-tuned on a subset of the Fruit Recognition dataset.

- **Developed by:** Mario Soler Vidal  
- **Model type:** Vision Transformer (ViT) for image classification  
- **Language(s):** Not applicable (Computer Vision)  
- **License:** MIT (or leave blank if unsure)  
- **Finetuned from model:** google/vit-base-patch16-224  

---

### Model Sources

- **Repository:** https://huggingface.co/Mariosolerzhawhugging/fruit-vit-model  
- **Demo:** https://huggingface.co/spaces/Mariosolerzhawhugging/fruit-classification-app  

---

## Uses

### Direct Use

This model can be used to classify images of fruits into one of the six supported categories. It is suitable for:

- Educational demonstrations  
- Image classification tasks  
- Computer vision experiments  

### Downstream Use

The model can be integrated into applications such as:

- Web apps (e.g., Gradio, Streamlit)  
- Retail or inventory systems  
- Automated fruit recognition pipelines  

### Out-of-Scope Use

- Classification of fruits outside the six trained classes  
- Complex real-world environments with heavy occlusion  
- Non-fruit image classification  

---

## Bias, Risks, and Limitations

The model was trained on images with mostly **clean and controlled backgrounds**, which may introduce bias.

Potential limitations:

- Reduced performance in highly cluttered or noisy environments  
- Limited generalization to unseen fruit types  
- Sensitivity to extreme lighting or image distortions  

### Recommendations

- Use the model primarily for the six trained fruit classes  
- Validate performance in real-world scenarios before deployment  
- Consider further fine-tuning for more diverse datasets  

---

## Training Details

### Training Data

The model was trained on a subset of the **Fruit Recognition dataset**, which contains over **44,000 images**.

Key characteristics:

- Images captured in a controlled lab environment  
- Resolution: 320 × 258 pixels  
- Mostly clean backgrounds  
- Variations in lighting, shadows, and pose  

---

### Training Procedure

#### Preprocessing

- Images resized to 224 × 224 pixels  
- Normalized using ImageNet statistics  
- Converted to RGB format  

Data augmentation:

- Random horizontal flip  
- Rotation  
- Color jitter  

---

#### Training Hyperparameters

- **Training regime:** fp32  
- **Approach:** Transfer learning (frozen backbone, trained classification head)  

---

## Evaluation

### Testing Data, Factors & Metrics

#### Testing Data

Validation and test splits from the Fruit Recognition dataset.

#### Metrics

- Accuracy  
- Macro F1 Score  

---

### Results

- **Accuracy:** ~0.999  
- **Macro F1 Score:** ~0.999  

#### Summary

The model achieves very high performance with confident predictions across all classes.

---

## Technical Specifications

### Model Architecture and Objective

- Vision Transformer (ViT)  
- Image classification objective  

### Compute Infrastructure

#### Hardware

GPU (training environment)

#### Software

- Python  
- Hugging Face Transformers  
- PyTorch  

---

## Model Card Authors

Mario Soler Vidal