Upload app.py with huggingface_hub

app.py

@@ -23,11 +23,6 @@ GPT_MAX_COMPLETION_TOKENS = 4096
 GPT_REASONING_EFFORT = "none"
 NOTES_PROMPT = "Transcribe the musical notes in this image. Return only the transcription."
 MAX_PREPROCESSED_SIDE = 1800
-TREBLE_TEMPLATE_PATHS = [
-    "examples/000100005-1_1_1.png",
-    "examples/000100016-1_3_1.png",
-    "examples/000100059-1_1_1.png",
-]
 
 CAMERA_CAPTURE_JS = """
 function () {
@@ -125,116 +120,17 @@ def _score_training_orientation(image: Image.Image):
     if width == 0 or height == 0:
         return -1
 
-    dark = gray < 235
-    aspect_score = width / max(height, 1)
-    horizontal_score = 0.0
-    if width > 40 and height > 20:
-        thresholded = (dark.astype("uint8") * 255)
-        horizontal_kernel = cv2.getStructuringElement(
-            cv2.MORPH_RECT,
-            (max(width // 20, 25), 1),
-        )
-        vertical_kernel = cv2.getStructuringElement(
-            cv2.MORPH_RECT,
-            (1, max(height // 8, 8)),
-        )
-        horizontal = cv2.morphologyEx(thresholded, cv2.MORPH_OPEN, horizontal_kernel)
-        vertical = cv2.morphologyEx(thresholded, cv2.MORPH_OPEN, vertical_kernel)
-        horizontal_score = float(horizontal.mean() - vertical.mean()) / 255.0
-
-    left_band = dark[:, :max(width // 5, 1)]
-    right_band = dark[:, -max(width // 5, 1):]
-    top_band = dark[:max(height // 2, 1), :]
-    bottom_band = dark[-max(height // 2, 1):, :]
-    left_density = float(left_band.mean())
-    right_density = float(right_band.mean())
-    top_density = float(top_band.mean())
-    bottom_density = float(bottom_band.mean())
-    clef_score = _score_treble_clef_left(gray)
-
-    # Training images read left-to-right: the clef/key signature should produce
-    # extra ink near the left edge, and the staff should be a wide horizontal strip.
-    return (
-        aspect_score * 3.0
-        + horizontal_score * 8.0
-        + (left_density - right_density) * 12.0
-        + (top_density - bottom_density) * 1.5
-        + clef_score * 20.0
-    )
-
-
-def _score_treble_clef_left(gray):
-    try:
-        import cv2
-        import numpy as np
-    except ImportError:
-        return 0.0
-
-    templates = _load_treble_templates()
-    if not templates:
-        return 0.0
-
-    height, width = gray.shape[:2]
-    search = gray[:, :max(width // 3, 1)]
-    search = cv2.resize(search, (max(search.shape[1], 60), max(search.shape[0], 60)))
-    search = 255 - search
-    best = 0.0
-
-    for template in templates:
-        for scale in (0.7, 0.9, 1.1, 1.3):
-            target_height = max(int(height * 0.7 * scale), 24)
-            aspect = template.shape[1] / max(template.shape[0], 1)
-            target_width = max(int(target_height * aspect), 12)
-            if target_height >= search.shape[0] or target_width >= search.shape[1]:
-                continue
-            resized = cv2.resize(template, (target_width, target_height))
-            result = cv2.matchTemplate(search, resized, cv2.TM_CCOEFF_NORMED)
-            if result.size:
-                best = max(best, float(np.max(result)))
-
-    return max(best, 0.0)
-
-
-_TREBLE_TEMPLATES = None
-
-
-def _load_treble_templates():
-    global _TREBLE_TEMPLATES
-    if _TREBLE_TEMPLATES is not None:
-        return _TREBLE_TEMPLATES
-
-    try:
-        import cv2
-        import numpy as np
-    except ImportError:
-        _TREBLE_TEMPLATES = []
-        return _TREBLE_TEMPLATES
-
-    templates = []
-    for path in TREBLE_TEMPLATE_PATHS:
-        if not os.path.exists(path):
-            continue
-        gray = np.array(Image.open(path).convert("L"))
-        height, width = gray.shape[:2]
-        left_region = gray[:, :max(width // 5, 1)]
-        ink = left_region < 235
-        rows = np.where(ink.any(axis=1))[0]
-        cols = np.where(ink.any(axis=0))[0]
-        if len(rows) == 0 or len(cols) == 0:
-            continue
-
-        top = max(int(rows[0]) - 8, 0)
-        bottom = min(int(rows[-1]) + 8, height - 1)
-        left = max(int(cols[0]) - 4, 0)
-        right = min(int(cols[-1]) + 12, left_region.shape[1] - 1)
-        template = left_region[top:bottom + 1, left:right + 1]
-        if template.shape[0] < 20 or template.shape[1] < 8:
-            continue
-
-        templates.append(255 - cv2.GaussianBlur(template, (3, 3), 0))
-
-    _TREBLE_TEMPLATES = templates
-    return _TREBLE_TEMPLATES
+    # Normalize out uneven camera lighting before comparing densities.
+    blur_size = min(max(min(width, height) // 6, 15) | 1, 201)
+    background = cv2.medianBlur(gray, blur_size)
+    normalized = cv2.divide(gray, background, scale=255)
+    ink = normalized < 128
+
+    band = max(width // 5, 1)
+    left_density = float(ink[:, :band].mean())
+    right_density = float(ink[:, -band:].mean())
+    # Treble clef + key sig always sit on the left, making that band denser.
+    return (width / max(height, 1)) * 2.0 + (left_density - right_density) * 15.0
 
 
 def _normalize_to_training_orientation(image: Image.Image):