Lpr Ocr Costa Rica

--- license: cc-by-4.0 language: - es tags: - license-plate-recognition - ocr - alpr - costa-rica - latin-america - onnx - ctc - crnn library_name: onnxruntime pipeline_tag: image-to-text datasets: [] metrics: - accuracy model-index: - name: wink-lpr-ocr-cr results: - task: type: image-to-text name: License Plate OCR metrics: - type: accuracy value: 98.4 name: Plate Accuracy - type: accuracy value: 97.8 name: Character Accuracy --- <div align="center"> <img src="https://www.wink.co/img/winklogo.png" width="200" alt="WINK Streaming"> # WINK LPR — Costa Rica Plate OCR **98.4% plate accuracy** · 4.4MB ONNX · CTC architecture · Real-time inference [](https://www.wink.co) [](https://creativecommons.org/licenses/by/4.0/) [](https://onnx.ai/) </div> ## Overview A high-accuracy OCR model for reading Costa Rica license plates, developed by [WINK Streaming](https://www.wink.co) as part of a production LPR (License Plate Recognition) system. The model reads cropped plate images and outputs the plate text. It handles both daytime color and nighttime IR/grayscale camera feeds. ## Model Details | Property | Value | |----------|-------| | **Architecture** | CNN (5 conv blocks) + BiLSTM (128 hidden × 2 layers) + CTC decoder | | **Parameters** | 1,117,170 | | **Input** | `[batch, 96, 192, 1]` — grayscale, NHWC, uint8 (0-255) | | **Output** | `[batch, 48, 38]` — 48 CTC timesteps, 38 classes | | **Alphabet** | `0-9`, `A-Z`, `_` (pad) + CTC blank (index 37) | | **ONNX size** | 4.4 MB | | **ONNX opset** | 18 (IR version 8) | | **Inference** | ~5ms CPU, ~2ms GPU | ## Performance | Metric | Score | |--------|-------| | **Plate accuracy** (exact match) | **98.4%** | | **Character accuracy** | **97.8%** | | Validation set | 9,074 crops | | Training set | 81,666 crops from 9,727 labeled detections | Evaluated on held-out 10% split of production data from Axis & Hikvision cameras (2304×1296, color day / IR night). ## Supported Plate Formats - **Standard:** 3 letters + 3 digits (e.g. `AAP096`) - **Numeric:** All digits (e.g. `672625`) - **Government:** 2-3 letters + 2 digits - **Motorcycle:** All digits, 2-r...

Source context: 0 downloads · 0 likes · Pipeline image-to-text · Library onnxruntime · Repo xmichael/lpr-ocr-costa-rica