kr-plate-detector β€” ν•œκ΅­ μ°¨λŸ‰ 번호판 κ²€μΆœ (RTMDet-tiny)

ν•œκ΅­ μ°¨λŸ‰ λ²ˆν˜ΈνŒμ„ κ²€μΆœν•˜λŠ” κ²½λŸ‰ λͺ¨λΈ. 100% ν•©μ„± λ°μ΄ν„°λ‘œλ§Œ ν•™μŠ΅ν–ˆμœΌλ©°, μ‹€μ œ ν•œκ΅­ 번호판(AI Hub 검증셋)μ—μ„œ μ„±λŠ₯을 μΈ‘μ •ν–ˆμŠ΅λ‹ˆλ‹€. ν”„λΌμ΄λ²„μ‹œ λΈ”λŸ¬ λ„κ΅¬μš©μœΌλ‘œ 개발.

  • μ•„ν‚€ν…μ²˜: RTMDet-tiny (mmdetection), 1클래슀(number plate)
  • μž…λ ₯: 960Γ—960, 좜λ ₯: 3-μŠ€μΌ€μΌ cls/bbox
  • λΌμ΄μ„ μŠ€: Apache-2.0 β€” 상업적 μ‚¬μš© κ°€λŠ₯
  • 라이브 데λͺ¨: https://huggingface.co/spaces/miraclenote78/privacy-blur

μ„±λŠ₯

평가 μ§€ν‘œ κ°’
ν•©μ„± 검증셋 (300μž₯) mAP50 / mAP50-95 0.949 / 0.806
싀사 ν•œκ΅­ 번호판 (AI Hub, 1,000μž₯) κ²€μΆœλ₯  (conf 0.25) 88.8%
〃 κ²€μΆœλ₯  (conf 0.40) 83.8%

싀사 검증은 AI Hub "μžλ™μ°¨ μ°¨μ’…/연식/번호판 μΈμ‹μš© μ˜μƒ"의 번호판 μ΄λ―Έμ§€λ‘œ μΈ‘μ •λ§Œ μˆ˜ν–‰(ν•™μŠ΅μ— λ―Έμ‚¬μš©). ν•©μ„±λ§ŒμœΌλ‘œ ν•™μŠ΅ν•œ λͺ¨λΈμ΄ μ‹€μ œ λ²ˆν˜ΈνŒμ— μΌλ°˜ν™”λ¨μ„ 확인.

μ‚¬μš©λ²•

ONNX Runtime (ꢌμž₯ β€” PyTorch/mmdet λΆˆν•„μš”)

from huggingface_hub import hf_hub_download
import cv2, numpy as np, onnxruntime as ort

onnx_path = hf_hub_download("miraclenote78/kr-plate-detector", "plate_kr_rtmdet_v4.onnx")
# μ „μ²˜λ¦¬Β·ν›„μ²˜λ¦¬λŠ” rtmdet_onnx.py μ°Έκ³  (동봉)
rt = hf_hub_download("miraclenote78/kr-plate-detector", "rtmdet_onnx.py")

import importlib.util
spec = importlib.util.spec_from_file_location("rtmdet_onnx", rt)
m = importlib.util.module_from_spec(spec); spec.loader.exec_module(m)

det = m.OnnxPlateDetector(onnx_path, conf_large=0.4, conf_tiny=0.4, tiny_frac=0.0)
img = cv2.imread("car.jpg")
for x1, y1, x2, y2, score in det.detect(img):
    cv2.rectangle(img, (int(x1), int(y1)), (int(x2), int(y2)), (0, 255, 0), 2)
cv2.imwrite("out.jpg", img)

μ „μ²˜λ¦¬: BGR keep-ratio 960 νŒ¨λ”© + mean/std μ •κ·œν™” Β· ν›„μ²˜λ¦¬: sigmoid β†’ distance λ””μ½”λ“œ β†’ NMS. 크기 쑰건뢀 λ¬Έν„± + κ°€λ‘œμ„Έλ‘œλΉ„ ν•„ν„°λ‘œ μ˜€κ²€μΆœ μ–΅μ œ.

ν•™μŠ΅ 방법 (μš”μ•½)

ν•œκ΅­ λ²ˆν˜ΈνŒμ€ κ·œκ²©Β·μ„œμ²΄κ°€ ν‘œμ€€ν™”λΌ μžˆμ–΄ ν”„λ‘œκ·Έλž¨μœΌλ‘œ ν•©μ„± 생성 κ°€λŠ₯: μ‹ ν˜• ν°νŒΒ·μ˜μ—…μš© λ…Έλž€νŒΒ·μ „κΈ°μ°¨ νŒŒλž€νŒΒ·κ΅¬ν˜• λ…Ήμƒ‰νŒμ„ κ·œμΉ™λŒ€λ‘œ λ Œλ”λ§ β†’ 배경에 원근 ν•©μ„± β†’ μ•Όκ°„/λͺ¨μ…˜λΈ”λŸ¬/λ…Έμ΄μ¦ˆ/JPEG 증강 β†’ YOLO 라벨 μžλ™ 생성. COCO μ‚¬μ „ν•™μŠ΅ RTMDet-tinyμ—μ„œ νŒŒμΈνŠœλ‹. (μƒμ„±κΈ°Β·ν•™μŠ΅ μ½”λ“œλŠ” λΉ„κ³΅κ°œ)

μ•Œλ €μ§„ ν•œκ³„

  • 싀사 검증은 번호판이 크게 λ³΄μ΄λŠ” 크둭 이미지 κΈ°μ€€. λ‹¬λ¦¬λŠ” 차의 원거리 μ†Œν˜• 번호판(λŒ€μ‹œμΊ  원경)은 별도 검증 ν•„μš” β€” μ„±λŠ₯이 λ‹€λ₯Ό 수 있음
  • κ΅¬ν˜• μ§€μ—­λͺ… μ„Έλ‘œμ“°κΈ° μ˜μ—…μš©νŒ, 극단적 κ°λ„λŠ” 약함
  • ν”„λΌμ΄λ²„μ‹œ μš©λ„ μ‹œ 완전성을 보μž₯ν•˜μ§€ μ•ŠμŒ β€” μ΅œμ’… κ²€μˆ˜λŠ” μ‚¬μš©μž μ±…μž„

좜처

ν”„λΌμ΄λ²„μ‹œ λΈ”λŸ¬ νŒŒμ΄ν”„λΌμΈμš©μœΌλ‘œ 개발. 데λͺ¨: privacy-blur Space.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support