Lab 5: Signature Detection & Cropper

Difficulty: Intermediate · Estimated time: ~3 hours

Objective

1. Load Grounding DINO Tiny model
2. Detect signature regions in scanned document images
3. Crop detected signatures with OpenCV
4. Output cropped PNGs + bounding box JSON
5. Serve via FastAPI endpoint

Step 1 — Install Grounding DINO

bash

pip install groundingdino-py
pip install opencv-python pillow fastapi uvicorn

Step 2 — Detect signatures

python

from groundingdino.util.inference import load_model, predict
import cv2
import json
from pathlib import Path

model = load_model("groundingdino_swint_ogc.pth", "GroundingDINO_SwinT_OGC.cfg.py")

def detect_signatures(image_path: str, threshold: float = 0.3):
    image = cv2.imread(image_path)
    boxes, logits, phrases = predict(
        model=model,
        image=image,
        caption="signature . handwritten signature",
        box_threshold=threshold,
        text_threshold=0.25,
    )
    results = []
    for box, score, phrase in zip(boxes, logits, phrases):
        h, w = image.shape[:2]
        x1, y1, x2, y2 = (box * [w, h, w, h]).astype(int)
        results.append({
            "bbox": [int(x1), int(y1), int(x2), int(y2)],
            "confidence": float(score),
            "label": phrase,
        })
    return results, image

def crop_signatures(image, detections: list, output_dir: str):
    Path(output_dir).mkdir(exist_ok=True)
    crops = []
    for i, det in enumerate(detections):
        x1, y1, x2, y2 = det["bbox"]
        crop = image[y1:y2, x1:x2]
        path = f"{output_dir}/signature_{i}.png"
        cv2.imwrite(path, crop)
        crops.append(path)
    return crops

Step 3 — FastAPI endpoint

python

from fastapi import FastAPI, UploadFile, File
import tempfile

app = FastAPI()

@app.post("/detect-signatures")
async def detect(file: UploadFile = File(...)):
    with tempfile.NamedTemporaryFile(suffix=".png") as tmp:
        tmp.write(await file.read())
        detections, image = detect_signatures(tmp.name)
        crops = crop_signatures(image, detections, "output")
    return {"detections": detections, "crops": crops}

Submission

GitHub repo with: detection script, FastAPI app, Dockerfile, sample results

Grading rubric

Criterion	Points
Grounding DINO loads and runs	20
Signatures detected with >70% recall	25
Cropped images saved correctly	20
FastAPI endpoint works	20
Bounding box JSON output	15
Total	100