README_ONNX.md

Hướng dẫn sử dụng DeepPrint ONNX Model

File ONNX model này chỉ chứa DeepPrint model (không có LoFTR). LoFTR là bước tiền xử lý riêng biệt.

File ONNX Model

• Đường dẫn: onnx_models/deepprint_fvc_db1_loftr_final.onnx
• Input: input_image - shape [batch_size, 1, 299, 299] (grayscale image)
• Outputs:
  • texture_embeddings: shape [batch_size, 256]
  • minutia_embeddings: shape [batch_size, 256]

Cài đặt

pip install onnxruntime

Cách sử dụng

1. Sử dụng script có sẵn

# Extract embedding từ 1 ảnh
python use_onnx_model.py --image1 path/to/fingerprint.tif

# So sánh 2 ảnh fingerprint
python use_onnx_model.py --image1 path/to/image1.tif --image2 path/to/image2.tif --threshold 0.5

2. Sử dụng trong code Python

import numpy as np
import onnxruntime as ort
from PIL import Image
import torchvision.transforms.functional as VTF
import sys

# Add DeepPrint path
sys.path.append("deepprint/fixed-length-fingerprint-extractors")
from flx.image_processing.binarization import LazilyAllocatedBinarizer
from flx.data.image_helpers import pad_and_resize_to_deepprint_input_size

# Load ONNX model
onnx_model_path = "onnx_models/deepprint_fvc_db1_loftr_final.onnx"
session = ort.InferenceSession(onnx_model_path)

# Preprocess image
def preprocess_image(image_path):
    img = Image.open(image_path).convert('L')
    img_tensor = VTF.to_tensor(img)
    
    # Binarize fingerprint
    binarizer = LazilyAllocatedBinarizer(5.0)
    img_tensor = binarizer(img_tensor)
    
    # Pad to square and resize to 299x299
    img_tensor = pad_and_resize_to_deepprint_input_size(img_tensor)
    img_tensor = img_tensor.unsqueeze(0)  # Add batch dimension
    
    return img_tensor.numpy().astype(np.float32)

# Extract embeddings
input_tensor = preprocess_image("path/to/fingerprint.tif")
outputs = session.run(['texture_embeddings', 'minutia_embeddings'], 
                      {'input_image': input_tensor})

texture_emb = outputs[0][0]  # [256]
minutia_emb = outputs[1][0]  # [256]
combined_emb = np.concatenate([texture_emb, minutia_emb])  # [512]

# Tính similarity giữa 2 embeddings
def cosine_similarity(emb1, emb2):
    emb1_norm = emb1 / (np.linalg.norm(emb1) + 1e-8)
    emb2_norm = emb2 / (np.linalg.norm(emb2) + 1e-8)
    return np.dot(emb1_norm, emb2_norm)

similarity = cosine_similarity(combined_emb1, combined_emb2)
print(f"Similarity: {similarity:.6f}")

Preprocessing Pipeline

Model yêu cầu preprocessing theo thứ tự sau:

1. Load image: PIL Image (grayscale)
2. Convert to tensor: VTF.to_tensor() → shape [1, H, W]
3. Binarize: LazilyAllocatedBinarizer(5.0) → enhance fingerprint ridges
4. Pad & Resize: pad_and_resize_to_deepprint_input_size() → [1, 299, 299]
5. Add batch dimension: unsqueeze(0) → [1, 1, 299, 299]