runtime/plugins/license_plate_recognition/manifest.toml

[plugin]
name = "license_plate_recognition"
version = "5.3.0"
display_name = "Nhận diện biển số xe"
description = "Phát hiện và nhận diện biển số xe Việt Nam"
library = "license_plate_recognition.dll"

[capabilities]
use_gpu = true
use_batching = true
use_tracking = true
use_zone = true

[classes]
names = ["car", "motorbike", "truck", "bus"]
defaults = ["car", "motorbike", "truck", "bus"]

[models]
# Order matters: the GUI's "Models" section binds `model_1` → files[0],
# `model_2` → files[1], etc. The order below MUST match the schema in
# `[[config.sections]]` labelled "MODELS" further down. Adding a new
# entry here requires bumping the schema in lockstep.
files = [
  "mbfs_vehicle_det_v4.onnx",            # model_1 — Detection (vehicle YOLO)
  "lp_detection_v4.onnx",                # model_2 — Plate Detector
  "mbfsn_cls_brand_car_v1.onnx",         # model_3 — Joint Make_Model classifier (80 classes, single softmax)
  "license_place_rec_crnn_v8.onnx",      # model_4 — CRNN/CTC sequence OCR (v8, TBC layout, handles 1+2 line via LPTR-AFLNet preprocessing)
  "weight_detect_corner.onnx",           # model_5 — 4-corner plate detector (perspective rectification)
]

[inference]
model = "mbfs_vehicle_det_v4.onnx"
confidence = 0.5
nms = 0.45
input_size = 640

[models.plate_detector]
model = "lp_detection_v4.onnx"
kind = "yolo"
input_size = 640
confidence = 0.3
nms = 0.45

# Joint Make_Model classifier (YOLOv8n-cls, 80 classes "{Make}_{Model}").
# Single softmax head; plugins split label on the first '_' to recover
# brand vs model. Replaces the older two-head model (brand_logits 9 +
# model_logits 72) — same datalake schema preserved.
[models.vehicle_brand]
model = "mbfsn_cls_brand_car_v1.onnx"
kind = "classifier"
input_size = 224

# CRNN/CTC sequence OCR (v8). Single model for both 1-line and 2-line
# Vietnamese plates.
#
# 1-line: corner-warp the plate crop to 128×32, feed to CRNN.
# 2-line: corner-warp to 128×64, split mid-row, resize each half to
#         64×32, horizontally concatenate to 128×32 (LPTR-AFLNet
#         preprocessing — arXiv 2507.16362), feed to CRNN.
#
# v8 ONNX output layout is TBC = [T=31, B, C=32]; the host's ocr_seq
# decoder auto-detects layout (BTC vs TBC) via shape introspection.
# Class 0 is the CTC blank; classes 1..=31 map to chars in `vocab`.
[models.plate_ocr_crnn]
model = "license_place_rec_crnn_v8.onnx"
kind = "ocr_seq"
input_size = [128, 32]   # [width, height]

[models.plate_ocr_crnn.ocr_seq]
vocab = "0123456789ABCDEFGHKLMNPRSTUVXYZ"
channels = 1   # grayscale; host applies BT.601 luma + x/255.0

# 4-corner plate corner detector. YOLO classifier with 4 classes
# (0=top_left, 1=top_right, 2=bottom_left, 3=bottom_right). The plugin
# picks the highest-confidence box per class, interpolates the 4th
# corner via the parallelogram rule when only 3 are present, then
# calls `host.warp_perspective` to produce an axis-aligned canvas
# before CRNN OCR. Falls back to the row_split.rs algorithm when
# fewer than 3 corners are detected.
[models.plate_corner]
model = "weight_detect_corner.onnx"
kind = "yolo"
input_size = 320
confidence = 0.3
nms = 0.5

# ── Config page schema (data-driven GUI) ──────────────────────────────────

[[config.sections]]
title = "MODELS"
fields = [
  { key = "model_1", type = "model", label = "Vehicle Detector:" },
  { key = "model_2", type = "model", label = "Plate Detector:" },
  { key = "model_3", type = "model", label = "Vehicle Brand:" },
  { key = "model_4", type = "model", label = "CRNN OCR Model (v8):" },
  { key = "model_5", type = "model", label = "Corner Detector:" },
]

[[config.sections]]
title = "DETECTION"
fields = [
  { key = "inference.confidence", type = "slider", label = "Vehicle Detection Confidence", help = "Primary YOLO vehicle detection confidence (0.0-1.0)", min = 0.0, max = 1.0, step = 0.05, default = 0.5 },
  { key = "models.plate_detector.confidence", type = "slider", label = "Plate Detection Confidence", help = "Plate detector confidence threshold (0.0-1.0)", min = 0.0, max = 1.0, step = 0.05, default = 0.3 },
  { key = "two_line_aspect_max", type = "slider", label = "2-Line Plate Aspect Threshold", help = "Plates with width/height below this are tried as 2-line first. Falls back to 1-line on failed validation. Lower = favor 1-line; higher = favor 2-line.", min = 1.5, max = 3.5, step = 0.1, default = 2.3 },
  { key = "min_ocr_confidence", type = "slider", label = "Min Plate OCR Confidence", help = "Reject whole plate readings below this averaged per-character confidence (0.0-1.0)", min = 0.0, max = 1.0, step = 0.05, default = 0.3 },
]

[[config.sections]]
title = "VEHICLE IDENTIFICATION"
fields = [
  { key = "vehicle_brand_enabled", type = "toggle", label = "Brand/Model Classifier", description = "Classify make and model on confirmed vehicle tracks", default = true },
  { key = "vehicle_brand_classify_interval", type = "number", label = "Classify Interval", help = "Frames between brand+model classification (1-30)", min = 1, max = 30, step = 1, default = 5, visible_when = "vehicle_brand_enabled" },
  { key = "vehicle_brand_min_confidence", type = "slider", label = "Min Brand Confidence", help = "Drop brand+model predictions below this joint-softmax confidence (80-class Make_Model head). 0.3 is a reasonable default; both brand and model receive the same confidence.", min = 0.0, max = 1.0, step = 0.05, default = 0.3, visible_when = "vehicle_brand_enabled" },
  { key = "vehicle_model_min_confidence", type = "slider", label = "Min Model Confidence", help = "Legacy knob from the two-head classifier era. The new joint Make_Model classifier uses vehicle_brand_min_confidence only; this slider is currently inert.", min = 0.0, max = 1.0, step = 0.05, default = 0.1, visible_when = "vehicle_brand_enabled" },
]

[[config.sections]]
title = "TRACKING"
fields = [
  { key = "tracker_min_observations", type = "number", label = "Min Observations", help = "Plate tracker observations before emitting result (1-10)", min = 1, max = 10, step = 1, default = 2 },
  { key = "dedup_cooldown_secs", type = "float", label = "Dedup Cooldown", help = "Seconds before same plate re-triggers (1-120)", min = 1.0, max = 120.0, step = 1.0, suffix = "s", default = 10.0 },
]