How we0091234/chinese_license_plate_detection_recognition works
Main Components
• Detection Module:
– Core detection scripts (e.g., detect_plate.py, detect_demo.py)
– Model definitions are stored in the “models” directory (various YOLO variants such as yolov5s, yolov5m, etc.)
– Data configuration and hyperparameter YAML files in the “data” folder (coco.yaml, voc.yaml, hyp configurations, etc.)
• Recognition Module:
– Plate recognition logic within the “plate_recognition” directory (scripts like plateNet.py, plate_rec.py, and merging/splitting utilities)
– This module likely fetches plate regions from the detection module and then applies a dedicated CNN or CRNN-based recognition process.
• Data & Configuration:
– “data” directory holds dataset configurations, training scripts (e.g., train2yolo.py, train.py), and data download scripts (in “data/scripts”).
– “weights” directory contains pre-trained model files that the detection and recognition modules load.
• Inference & Deployment:
– Additional inference scripts (onnx_infer.py, openvino_infer.py) support deploying models using ONNX, TensorRT (see torch2trt folder), and OpenVINO.
– The project also provides deployment scripts and even links to Android/NCNN examples.
• Utilities & Helper Modules:
– “utils” directory includes data loaders, loss functions, metrics, plotting utilities, and modules for integration with external systems (e.g., AWS, Google App Engine).
– “torch2trt” folder offers conversion utilities for TensorRT optimization.
Relationships and Interactions
• Data Flow:
– Input images/videos (from the “imgs” folder for demos) are processed by the Detection Module.
– The detected license plate areas or regions of interest are passed to the Recognition Module.
– Processed output (detected and recognized plate numbers) is saved to an “output” or “result” directory.
• Model Training & Tuning:
– Data configuration files (YAMLs within the “data” folder) enable training both detection and recognition networks.
– Training scripts like train.py, ccpt_process.py, and test scripts coordinate training/validation benchmarks.
• Deployment Paths:
– Separate inference pipelines exist for different model formats (pure PyTorch, ONNX, TensorRT, OpenVINO).
– Conversion/inference scripts show how the same core detection-recognition pipeline is repurposed for production deployments.
Architectural Patterns and Design Principles
• Modular Design:
– The project separates concerns by having dedicated directories and scripts for detection, recognition, training, inference, and utility functions.
• Configuration-Driven:
– Extensive use of YAML configuration files for datasets, model hyperparameters, and training policies.
• Multi-Deployment Strategy:
– Flexibility is provided by having different inference scripts for various deployment targets (e.g., onnx_infer.py, openvino_infer.py, Torch2TRT tools).
• Extensibility:
– The separation of detection and recognition facilitates independent updates or swaps for each module.
Key Technologies and Frameworks
• Python as the primary language.
• PyTorch for model training and inference.
• YOLO-based model architectures for detection.
• Deployment tools: ONNX, TensorRT, OpenVINO for converting and serving the models.
• Supplementary libraries for data augmentation, loss computations, metrics, logging (including wandb_logging).
Putting It All Together
• Branch out from that block into at least four high-level modules:
– Input Processing (images, videos)
– Detection Module (and details: model configuration, YOLO variants, data YAMLs)
– Recognition Module (plate recognition scripts, plateNet, plate_rec)
– Utility/Training/Deployment (training scripts, conversion tools, and integration utilities)
• Interconnections, such as configuration flows (YAML files feeding into training and detection modules) and model weight usage.