File size: 2,974 Bytes
dce1dbf | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 | ---
license: apache-2.0
library_name: keras
pipeline_tag: image-classification
tags:
- emotion-detection
- facial-expression-recognition
- image-classification
- keras
- tensorflow
- efficientnetv2
---
# Facial Emotion Recognition — Model
Detects the emotion expressed in a photo of a human face.
Facial emotion classification model in Keras 3 format (`best_model.keras`).
Given a cropped face image it outputs probabilities over 7 emotion classes.
- **Maintainer of this repository**: Mamd1
- **Repository**: `Mamd1/emotion-detection-model`
- **Demo Space**: `Mamd1/emotion-detection-app`
- **License**: Apache-2.0
## Attribution
These weights were **not trained by the maintainer of this repository**. They are
redistributed from the original Apache-2.0 licensed release by **Ali Ahmed**
(`ali-ahmed-ai-developer/emotion_detection_model` on the Hugging Face Hub).
Apache-2.0 permits redistribution and modification provided the license and
attribution are preserved — this notice satisfies that requirement. Do not
remove it.
## Classes
Index order matters — the output vector is in exactly this order:
| Index | Class |
|-------|----------|
| 0 | angry |
| 1 | disgust |
| 2 | fear |
| 3 | happy |
| 4 | neutral |
| 5 | sad |
| 6 | surprise |
## Model details
- **Backbone**: EfficientNetV2-S (ImageNet pre-trained)
- **Head**: GlobalAveragePooling2D → BatchNorm → Dropout → Dense(512) → BatchNorm → Dropout → Dense(7)
- **Input shape**: `(380, 380, 3)`, float32, values in `[0, 255]` (EfficientNetV2 rescales internally)
- **Parameters**: ~21M
- **File**: `best_model.keras` (~343 MB, stored via Git LFS)
## Preprocessing
The reference application converts the image to grayscale, then back to RGB
(so R=G=B across 3 channels), resizes to 380x380, and passes raw `[0, 255]`
float32 values. Matching this preprocessing is required for accurate results.
## Usage
```python
import numpy as np
import tensorflow as tf
from PIL import Image
from huggingface_hub import hf_hub_download
CLASSES = ['angry', 'disgust', 'fear', 'happy', 'neutral', 'sad', 'surprise']
path = hf_hub_download(
repo_id="Mamd1/emotion-detection-model",
filename="best_model.keras",
)
model = tf.keras.models.load_model(path)
image = Image.open("face.jpg").convert("L").convert("RGB").resize((380, 380))
x = np.expand_dims(np.array(image, dtype=np.float32), axis=0)
probs = model.predict(x)[0]
print(dict(zip(CLASSES, probs.tolist())))
```
## Limitations
- Trained on tightly cropped faces. Wide shots, small faces, or busy backgrounds
degrade accuracy significantly — crop to the face first.
- Facial expression recognition is not a reliable measure of a person's actual
internal emotional state. Do not use for consequential decisions about people
(hiring, security screening, medical or psychological assessment).
- Accuracy varies across demographic groups and lighting conditions.
|