ANPR/config.py at main · VedantGupta-DTU/ANPR · GitHub

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
"""
Configuration settings for License Plate Recognition Pipeline
"""
import os

# Performance mode: disabled — full accuracy pipeline with consensus voting
FAST_MODE = False

# Paths
BASE_DIR = os.path.dirname(os.path.abspath(__file__))
MODEL_PATH = os.path.join(BASE_DIR, "best.pt")
MODEL_PATH_ONNX = os.path.join(BASE_DIR, "best.onnx")
MODEL_PATH_COREML = os.path.join(BASE_DIR, "best.mlpackage")
OUTPUT_DIR = os.path.join(BASE_DIR, "output")
TEST_IMAGES_DIR = os.path.join(BASE_DIR, "test_images")

# YOLO inference input size
YOLO_IMGSZ = 640

# Create directories if they don't exist
os.makedirs(OUTPUT_DIR, exist_ok=True)
os.makedirs(TEST_IMAGES_DIR, exist_ok=True)

# YOLO Detection settings
DETECTION_CONFIDENCE = 0.50  # Balanced — catches plates without picking up signboards
IOU_THRESHOLD = 0.45         # IoU threshold for NMS

# Bike / small plate settings
SMALL_PLATE_AREA_THRESHOLD = 15000  # px² – plates below this get extra padding
SMALL_PLATE_PADDING = 25            # Extra pixels around small/bike plates
IGNORE_BIKE_PLATES = False           # Skip square/tall plates (bikes) and only detect cars

# DeepSeek OCR-2 settings
OCR_MODEL_NAME = "deepseek-ai/DeepSeek-OCR-2"
OCR_PROMPT = "<image>\nFree OCR. "  # Prompt for clean text extraction
USE_FLASH_ATTENTION = True  # Set to False if flash-attn not installed

# Processing settings
DEVICE = "cuda"  # Use "cpu" if no GPU available
IMAGE_EXTENSIONS = [".jpg", ".jpeg", ".png", ".bmp", ".webp"]

# Video settings
VIDEO_EXTENSIONS = [".mp4", ".avi", ".mov", ".mkv"]
VIDEO_FRAME_SKIP = 1  # Process every frame for live — don't miss cars

# Tracking / confirmation settings (video_pipeline.py)
# Only output a plate if its track is stable across multiple sampled frames.
VIDEO_MIN_TRACK_READS = 1

# Require the best plate text to dominate OCR reads for that track.
# This helps avoid “valid but noisy” OCR where different valid plates appear
# across frames for the same vehicle.
# Set to 0 to effectively disable the stability gate.
VIDEO_MIN_PLATE_STABILITY_RATIO = 0.0  # best_key_freq / pool_len

# Filter confirmed tracks by YOLO detection confidence (reduces false positives).
VIDEO_MIN_DET_CONF = 0.3

# For each plate track, only vote using the last portion of reads (later
# frames are usually less blurry because the car is closer).
VIDEO_OCR_TAIL_RATIO = 0.6

# Prefer truly confident OCR when multiple valid-looking plates compete.
# (PaddleOCR rec_scores are typically in [0,1].)
VIDEO_MIN_OCR_CONF = 0.45

# OCR confidence weighting power in track voting.
VIDEO_OCR_CONF_POWER = 3.0
VIDEO_LATENCY_LOG = os.path.join(OUTPUT_DIR, "latency_report.json")

# Edge / Raspberry Pi settings
MAX_PENDING_OCR = 5  # Drop new OCR tasks when queue exceeds this (prevents lag on Pi)

# ──────────────────────────────────────────────────────────────────
# Night Mode — auto-applied when dark scenes are detected
# ──────────────────────────────────────────────────────────────────
NIGHT_BRIGHTNESS_THRESHOLD = 80    # mean(gray) below this → night mode
NIGHT_DARK_RATIO_THRESHOLD = 0.45  # fraction of pixels < 60 for night
NIGHT_DETECTION_CONFIDENCE = 0.25  # much lower — let temporal voting filter
NIGHT_MIN_DET_CONF = 0.20          # accept dimmer detections
NIGHT_MIN_OCR_CONF = 0.30          # accept noisier OCR
NIGHT_FRAME_SKIP = 1               # process every frame at night
NIGHT_CLAHE_CLIP = 5.0             # aggressive contrast boost
NIGHT_CHECK_INTERVAL = 10          # re-evaluate day/night every N frames