🤟 ASL Sign Language Recognition — EfficientNet-B0
A fine-tuned EfficientNet-B0 model for recognizing American Sign Language (ASL) alphabet letters from images. Achieves 99.02% accuracy on the evaluation set.
Model Details
| Property | Value |
|---|---|
| Base Model | google/efficientnet-b0 (ImageNet pretrained) |
| Parameters | 4,040,854 (~15.6MB) |
| Input Size | 224×224 RGB images |
| Classes | 26 (A-Z ASL alphabet letters) |
| Inference Speed | <10ms/frame on GPU, ~30ms on CPU |
Training
| Hyperparameter | Value |
|---|---|
| Learning Rate | 2e-4 |
| Batch Size | 16 |
| Epochs | 5 |
| Optimizer | AdamW |
| LR Scheduler | Cosine |
| Weight Decay | 1e-4 |
| Warmup Ratio | 5% |
Training Results
| Epoch | Eval Accuracy | Eval Loss |
|---|---|---|
| 1 | 89.45% | 0.405 |
| 2 | 97.67% | 0.096 |
| 3 | 98.28% | 0.056 |
| 4 | 98.71% | 0.047 |
| 5 | 99.02% | 0.036 |
Data Augmentation
- RandomResizedCrop (scale 0.8-1.0)
- RandomHorizontalFlip (p=0.3)
- RandomRotation (±15°)
- ColorJitter (brightness=0.3, contrast=0.3, saturation=0.2, hue=0.1)
Usage
from transformers import pipeline
classifier = pipeline("image-classification", model="abdollahhh/asl-sign-language-efficientnet-b0")
result = classifier("path/to/hand_sign.jpg")
print(result)
# [{'label': 'A', 'score': 0.98}, ...]
Manual inference
from transformers import AutoImageProcessor, AutoModelForImageClassification
from PIL import Image
import torch
processor = AutoImageProcessor.from_pretrained("abdollahhh/asl-sign-language-efficientnet-b0")
model = AutoModelForImageClassification.from_pretrained("abdollahhh/asl-sign-language-efficientnet-b0")
model.eval()
image = Image.open("hand_sign.jpg")
inputs = processor(images=image, return_tensors="pt")
with torch.no_grad():
logits = model(**inputs).logits
predicted_class = logits.argmax(-1).item()
label = model.config.id2label[str(predicted_class)]
print(f"Predicted: {label}")
Live Demo
Try the real-time webcam demo: ASL Sign Language Recognition Space
Dataset
Trained on Marxulia/asl_sign_languages_alphabets_v03:
- 10,873 images total (9,242 train / 1,631 eval)
- 26 classes: A through Z
- Stratified 85/15 train/eval split
Limitations
- Trained on controlled studio images — may have reduced accuracy with varied backgrounds/lighting
- Only recognizes static letter signs (A-Z), not dynamic gestures (J, Z involve motion)
- Works best with a clean hand against a neutral background