feat: ocr (#18836)

* feat: add OCR functionality and related configurations

* chore: update labeler configuration for machine learning files

* feat(i18n): enhance OCR model descriptions and add orientation classification and unwarping features

* chore: update Dockerfile to include ccache for improved build performance

* feat(ocr): enhance OCR model configuration with orientation classification and unwarping options, update PaddleOCR integration, and improve response structure

* refactor(ocr): remove OCR_CLEANUP job from enum and type definitions

* refactor(ocr): remove obsolete OCR entity and migration files, and update asset job status and schema to accommodate new OCR table structure

* refactor(ocr): update OCR schema and response structure to use individual coordinates instead of bounding box, and adjust related service and repository files

* feat: enhance OCR configuration and functionality

- Updated OCR settings to include minimum detection box score, minimum detection score, and minimum recognition score.
- Refactored PaddleOCRecognizer to utilize new scoring parameters.
- Introduced new database tables for asset OCR data and search functionality.
- Modified related services and repositories to support the new OCR features.
- Updated translations for improved clarity in settings UI.

* sql changes

* use rapidocr

* change dto

* update web

* update lock

* update api

* store positions as normalized floats

* match column order in db

* update admin ui settings descriptions

fix max resolution key

set min threshold to 0.1

fix bind

* apply config correctly, adjust defaults

* unnecessary model type

* unnecessary sources

* fix(ocr): switch RapidOCR lang type from LangDet to LangRec

* fix(ocr): expose lang_type (LangRec.CH) and font_path on OcrOptions for RapidOCR

* fix(ocr): make OCR text search case- and accent-insensitive using ILIKE + unaccent

* fix(ocr): add OCR search fields

* fix: Add OCR database migration and update ML prediction logic.

* trigrams are already case insensitive

* add tests

* format

* update migrations

* wrong uuid function

* linting

* maybe fix medium tests

* formatting

* fix weblate check

* openapi

* sql

* minor fixes

* maybe fix medium tests part 2

* passing medium tests

* format web

* readd sql

* format dart

* disabled in e2e

* chore: translation ordering

---------

Co-authored-by: mertalev <101130780+mertalev@users.noreply.github.com>
Co-authored-by: Alex Tran <alex.tran1502@gmail.com>

machine-learning/immich_ml/models/ocr/detection.py

@@ -0,0 +1,86 @@
+from typing import Any
+
+import numpy as np
+from PIL import Image
+from rapidocr.ch_ppocr_det import TextDetector as RapidTextDetector
+from rapidocr.inference_engine.base import FileInfo, InferSession
+from rapidocr.utils import DownloadFile, DownloadFileInput
+from rapidocr.utils.typings import EngineType, LangDet, OCRVersion, TaskType
+from rapidocr.utils.typings import ModelType as RapidModelType
+
+from immich_ml.config import log
+from immich_ml.models.base import InferenceModel
+from immich_ml.models.transforms import decode_cv2
+from immich_ml.schemas import ModelFormat, ModelSession, ModelTask, ModelType
+from immich_ml.sessions.ort import OrtSession
+
+from .schemas import OcrOptions, TextDetectionOutput
+
+
+class TextDetector(InferenceModel):
+    depends = []
+    identity = (ModelType.DETECTION, ModelTask.OCR)
+
+    def __init__(self, model_name: str, **model_kwargs: Any) -> None:
+        super().__init__(model_name, **model_kwargs, model_format=ModelFormat.ONNX)
+        self.max_resolution = 736
+        self.min_score = 0.5
+        self.score_mode = "fast"
+        self._empty: TextDetectionOutput = {
+            "image": np.empty(0, dtype=np.float32),
+            "boxes": np.empty(0, dtype=np.float32),
+            "scores": np.empty(0, dtype=np.float32),
+        }
+
+    def _download(self) -> None:
+        model_info = InferSession.get_model_url(
+            FileInfo(
+                engine_type=EngineType.ONNXRUNTIME,
+                ocr_version=OCRVersion.PPOCRV5,
+                task_type=TaskType.DET,
+                lang_type=LangDet.CH,
+                model_type=RapidModelType.MOBILE if "mobile" in self.model_name else RapidModelType.SERVER,
+            )
+        )
+        download_params = DownloadFileInput(
+            file_url=model_info["model_dir"],
+            sha256=model_info["SHA256"],
+            save_path=self.model_path,
+            logger=log,
+        )
+        DownloadFile.run(download_params)
+
+    def _load(self) -> ModelSession:
+        # TODO: support other runtime sessions
+        session = OrtSession(self.model_path)
+        self.model = RapidTextDetector(
+            OcrOptions(
+                session=session.session,
+                limit_side_len=self.max_resolution,
+                limit_type="min",
+                box_thresh=self.min_score,
+                score_mode=self.score_mode,
+            )
+        )
+        return session
+
+    def _predict(self, inputs: bytes | Image.Image) -> TextDetectionOutput:
+        results = self.model(decode_cv2(inputs))
+        if results.boxes is None or results.scores is None or results.img is None:
+            return self._empty
+        return {
+            "image": results.img,
+            "boxes": np.array(results.boxes, dtype=np.float32),
+            "scores": np.array(results.scores, dtype=np.float32),
+        }
+
+    def configure(self, **kwargs: Any) -> None:
+        if (max_resolution := kwargs.get("maxResolution")) is not None:
+            self.max_resolution = max_resolution
+            self.model.limit_side_len = max_resolution
+        if (min_score := kwargs.get("minScore")) is not None:
+            self.min_score = min_score
+            self.model.postprocess_op.box_thresh = min_score
+        if (score_mode := kwargs.get("scoreMode")) is not None:
+            self.score_mode = score_mode
+            self.model.postprocess_op.score_mode = score_mode