Improve OCR speed and preprocessing

VRWT_Tool/VRC_OSC/.env.example

@@ -1,3 +1,11 @@
 DISCORD_WEBHOOK_URL=
 VRCHAT_USERNAME=
 VRCHAT_PASSWORD=
+
+OCR_CROP_LEFT=0.05
+OCR_CROP_TOP=0.35
+OCR_CROP_RIGHT=0.95
+OCR_CROP_BOTTOM=0.95
+OCR_PREPROCESS=1
+OCR_SCALE=1.5
+OCR_USE_ANGLE_CLS=0

VRWT_Tool/VRC_OSC/readme.txt

@@ -61,12 +61,39 @@ missing_count = 3
 DISCORD_WEBHOOK_URL=
 VRCHAT_USERNAME=
 VRCHAT_PASSWORD=
+
+OCR_CROP_LEFT=0.05
+OCR_CROP_TOP=0.35
+OCR_CROP_RIGHT=0.95
+OCR_CROP_BOTTOM=0.95
+OCR_PREPROCESS=1
+OCR_SCALE=1.5
+OCR_USE_ANGLE_CLS=0
 ```
 
 `DISCORD_WEBHOOK_URL` は VRC ログ結果の Discord Webhook 通知に使う。
 
 `VRCHAT_USERNAME` / `VRCHAT_PASSWORD` は `src/getUser.py` を使う場合だけ必要。
 
+OCR 設定は速度と精度の調整用。
+
+`OCR_CROP_*` は VRChat ウィンドウ内の相対座標。`0.0` から `1.0` で指定する。デフォルトは画面下側の中央寄りだけを読む。
+
+フル画面 OCR に戻す場合。
+
+```env
+OCR_CROP_LEFT=0
+OCR_CROP_TOP=0
+OCR_CROP_RIGHT=1
+OCR_CROP_BOTTOM=1
+```
+
+`OCR_PREPROCESS=1` はグレースケール化、コントラスト補正、シャープ化、拡大を行う。
+
+`OCR_SCALE` は前処理時の拡大率。大きいほど小さい文字に強くなるが遅くなる。
+
+`OCR_USE_ANGLE_CLS=0` は角度分類を切って高速化する。縦書きや回転文字を読む必要がある場合だけ `1` にする。
+
 ## OSC パラメータ
 
 Avatar Parameters 側の名前と一致させる。
@@ -123,6 +150,8 @@ notice 条件に入った場合。
 
 `translation` は OCR 結果を翻訳し、翻訳結果をクリップボードへコピーする。
 
+OCR はデフォルトで画面全体ではなく、画面下側の中央寄りを切り出して読む。会話ログや字幕を読む用途では速く、余計な UI 文字を拾いにくい。
+
 ## runtime
 
 新規の実行ログは基本的に以下へまとめる。

VRWT_Tool/VRC_OSC/src/common/env_config.py

@@ -0,0 +1,56 @@
+import os
+
+from common.project_paths import ENV_FILE
+
+
+def readDotEnvValue(key):
+    if not ENV_FILE.exists():
+        return ""
+
+    for line in ENV_FILE.read_text(encoding="utf-8-sig", errors="replace").splitlines():
+        stripped = line.strip()
+        if not stripped or stripped.startswith("#"):
+            continue
+        if not stripped.startswith(key + "="):
+            continue
+
+        value = stripped.split("=", 1)[1].strip()
+        return value.strip('"').strip("'")
+
+    return ""
+
+
+def getEnvValue(key, default=""):
+    value = os.environ.get(key, "").strip()
+    if value:
+        return value
+
+    value = readDotEnvValue(key).strip()
+    if value:
+        return value
+
+    return default
+
+
+def getEnvBool(key, default=False):
+    value = getEnvValue(key, "").strip().lower()
+    if not value:
+        return default
+
+    if value in ("1", "true", "yes", "on"):
+        return True
+    if value in ("0", "false", "no", "off"):
+        return False
+
+    return default
+
+
+def getEnvFloat(key, default=0.0):
+    value = getEnvValue(key, "").strip()
+    if not value:
+        return default
+
+    try:
+        return float(value)
+    except ValueError:
+        return default

VRWT_Tool/VRC_OSC/src/ocr/ocr_actions.py

@@ -1,12 +1,6 @@
 import os
 from datetime import datetime
 
-import pyperclip
-from paddleocr import PaddleOCR
-
-from common.runtime_log import appendRuntimeLog
-from screen.screen_actions import captureOcrRegion
-
 os.environ["FLAGS_enable_pir_api"] = "0"
 os.environ["FLAGS_use_onednn"] = "0"
 os.environ["FLAGS_use_mkldnn"] = "0"
@@ -14,6 +8,17 @@ os.environ["FLAGS_use_onednn_bfloat16"] = "0"
 os.environ["MKLDNN_DISABLE_WORKSPACE"] = "1"
 os.environ["ONEDNN_VERBOSE"] = "0"
 
+import pyperclip
+from paddleocr import PaddleOCR
+from PIL import Image
+from PIL import ImageFilter
+from PIL import ImageOps
+
+from common.env_config import getEnvBool
+from common.env_config import getEnvFloat
+from common.runtime_log import appendRuntimeLog
+from screen.screen_actions import captureOcrRegion
+
 _ocr_engine = None
 
 
@@ -27,10 +32,17 @@ def getOcrEngine():
 
     if _ocr_engine is None:
         log("INFO", "PaddleOCR initializing")
-        _ocr_engine = PaddleOCR(
-            lang="japan",
-            use_angle_cls=True,
-        )
+        options = {
+            "lang": "japan",
+            "use_angle_cls": getEnvBool("OCR_USE_ANGLE_CLS", False),
+            "show_log": False,
+        }
+
+        try:
+            _ocr_engine = PaddleOCR(**options)
+        except (TypeError, ValueError):
+            options.pop("show_log", None)
+            _ocr_engine = PaddleOCR(**options)
 
     return _ocr_engine
 
@@ -99,6 +111,37 @@ def printOcrResult(text):
         print(line, flush=True)
 
 
+def preprocessOcrImage(image_path):
+    enabled = getEnvBool("OCR_PREPROCESS", True)
+    if not enabled:
+        return image_path
+
+    try:
+        image = Image.open(image_path)
+        image = image.convert("L")
+        image = ImageOps.autocontrast(image)
+        image = image.filter(ImageFilter.SHARPEN)
+
+        scale = max(1.0, min(2.0, getEnvFloat("OCR_SCALE", 1.5)))
+        if scale > 1.0:
+            width, height = image.size
+            resampling = getattr(
+                getattr(Image, "Resampling", Image),
+                "LANCZOS",
+            )
+            image = image.resize(
+                (int(width * scale), int(height * scale)),
+                resampling,
+            )
+
+        image.save(image_path)
+        log("INFO", f"OCR image preprocessed={image_path}")
+    except Exception as e:
+        log("ERROR", f"OCR preprocess failed detail={e}")
+
+    return image_path
+
+
 def runPaddleOcr(engine, image_path):
     if hasattr(engine, "predict"):
         return engine.predict(str(image_path))
@@ -109,6 +152,7 @@ def runOcrFromImage(image_path, show_result=True):
     log("ACTION", f"PaddleOCR start image={image_path}")
 
     try:
+        image_path = preprocessOcrImage(image_path)
         engine = getOcrEngine()
         result = runPaddleOcr(engine, image_path)
     except Exception as e:

VRWT_Tool/VRC_OSC/src/screen/screen_actions.py

@@ -1,6 +1,7 @@
 import pyautogui
 import pygetwindow as gw
 
+from common.env_config import getEnvFloat
 from common.project_paths import RUNTIME_DIR
 
 SCREENSHOT_PATH = RUNTIME_DIR / "latest_screenshot.png"
@@ -93,10 +94,44 @@ def getVrchatRegion():
     return left, top, width, height
 
 
-def captureVrchatWindowTo(image_path):
+def clampRatio(value):
+    return max(0.0, min(1.0, value))
+
+
+def getOcrRegion():
+    left, top, width, height = getVrchatRegion()
+
+    # Default to the lower central area. It is much faster than OCRing the
+    # whole VRChat window and usually contains chat/subtitle text.
+    crop_left = clampRatio(getEnvFloat("OCR_CROP_LEFT", 0.05))
+    crop_top = clampRatio(getEnvFloat("OCR_CROP_TOP", 0.35))
+    crop_right = clampRatio(getEnvFloat("OCR_CROP_RIGHT", 0.95))
+    crop_bottom = clampRatio(getEnvFloat("OCR_CROP_BOTTOM", 0.95))
+
+    if crop_right <= crop_left:
+        crop_left = 0.0
+        crop_right = 1.0
+
+    if crop_bottom <= crop_top:
+        crop_top = 0.0
+        crop_bottom = 1.0
+
+    region_left = left + int(width * crop_left)
+    region_top = top + int(height * crop_top)
+    region_width = max(1, int(width * (crop_right - crop_left)))
+    region_height = max(1, int(height * (crop_bottom - crop_top)))
+
+    log(
+        "INFO",
+        f"OCR crop left={region_left} top={region_top} width={region_width} height={region_height}",
+    )
+    return region_left, region_top, region_width, region_height
+
+
+def captureRegionTo(image_path, region):
     image_path.parent.mkdir(parents=True, exist_ok=True)
 
-    left, top, width, height = getVrchatRegion()
+    left, top, width, height = region
 
     image = pyautogui.screenshot(region=(left, top, width, height))
     image.save(image_path)
@@ -105,12 +140,22 @@ def captureVrchatWindowTo(image_path):
     return image_path
 
 
+def captureVrchatWindowTo(image_path):
+    return captureRegionTo(
+        image_path,
+        getVrchatRegion(),
+    )
+
+
 def captureVrchatWindow():
     return captureVrchatWindowTo(SCREENSHOT_PATH)
 
 
 def captureOcrRegion():
-    return captureVrchatWindowTo(OCR_SCREENSHOT_PATH)
+    return captureRegionTo(
+        OCR_SCREENSHOT_PATH,
+        getOcrRegion(),
+    )
 
 
 def captureScreen():