feat(python): add PP-OCRv6 paddle backend support

[jaminmei]

Adds PP-OCRv6 paddle backend support alongside the existing onnxruntime/openvino. Two changes were needed beyond just registering model URLs:

1. init_predictor + device_config: v6 paddle models use the PIR format (inference.json), same as v5. Added PPOCRV6 to the existing PPOCRV5 branches in both paddle/main.py and paddle/device_config.py so v6 gets the same predictor configuration (enable_new_ir, enable_new_executor, set_optimization_level).

2. Model registry: Added paddle.PP-OCRv6 section with det + rec (tiny/small/medium), pointing to the official files on RapidAI/RapidOCR master. rec entries include dict_url pointing to the dict files uploaded by the maintainer.

Verified: v6 paddle det/rec tests pass; v4/v5 paddle regression tests all pass (no behavior change).

[jaminmei]

Hi @SWHL, while working on the MNN backend for v6, I noticed that the v6 dict files haven't been uploaded to the RapidAI/RapidOCR ModelScope repo yet.

For v4 and v5, these dict files are hosted under paddle/PP-OCRvX/rec/<model_dir>/, and other backends that can't read embedded character info (MNN, torch) reference them via dict_url.

For v6, the dict files currently only exist in the PaddleOCR GitHub source repo: https://github.com/PaddlePaddle/PaddleOCR/tree/main/ppocr/utils/dict

• ppocrv6_dict.txt (18708 chars)
• ppocrv6_tiny_dict.txt (6904 chars)

I think the missing v6 dict files have caused some complications. On one hand, PR #696 required more adjustments to the engine code (have_key / get_character_list) than would otherwise be necessary. On the other hand, the MNN backend also needs the dict_url for rec models.

If the dict files are uploaded to paddle/PP-OCRv6/rec/ following the same structure as v4/v5, the engine workaround in PR #696 can be reverted, and all backends can reference them via the standard dict_url path without special handling.

[jaminmei]

@SWHL Since the dict files are now available on the repo, I've submitted a new commit that reverts the have_key / get_character_list workaround and adds dict_url to the paddle v6 rec entries. No engine code changes beyond the init_predictor and device_config version branches needed for PIR format support.

[Copilot]

Pull request overview

This PR extends the Python Paddle inference backend to support PP-OCRv6 models (in addition to existing onnxruntime/openvino support) and wires that support into the default model registry and tests.

Changes:

• Treat OCRVersion.PPOCRV6 the same as PPOCRV5 for Paddle predictor initialization (enable_memory_optim() path).
• Apply the same Paddle CPU “new IR / new executor / optimization level” configuration for PPOCRV6 as for PPOCRV5.
• Register Paddle PP-OCRv6 det/rec default model bundles and run PP-OCRv6 det/rec tests with the Paddle engine.

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 1 comment.

File	Description
python/tests/test_engine.py	Adds Paddle to the PP-OCRv6 det/rec parametrized engine coverage.
python/rapidocr/inference_engine/paddle/main.py	Extends Paddle predictor init logic to include PP-OCRv6.
python/rapidocr/inference_engine/paddle/device_config.py	Applies PPOCRv5-style CPU IR/executor settings to PPOCRv6.
python/rapidocr/default_models.yaml	Adds Paddle PP-OCRv6 det/rec model registry entries.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[jaminmei]

Thanks for catching this. The initial commit did implement the inference.yml approach (hence the PR description), but after the dict files were uploaded to the repo, a follow-up commit reverted that workaround and switched to the standard dict_url path. The current behavior is intentional: have_key() returns False, dict_url is present in the rec entries, and the recognizer downloads the dict via get_dict_key_url() as expected. I'll update the PR description to match.