mathcraft-ocr 0.2.3__tar.gz → 0.2.5__tar.gz

{mathcraft_ocr-0.2.3/mathcraft_ocr.egg-info → mathcraft_ocr-0.2.5}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: mathcraft-ocr
-Version: 0.2.3
+Version: 0.2.5
 Summary: ONNX-only OCR runtime for mathematical documents
 Author: SakuraMathcraft
 License-Expression: MIT
@@ -196,6 +196,10 @@ Model artifacts are downloaded from the MathCraft-Models release assets declared
 
 The actual provider is available on results through the `provider` field.
 
+## Benchmarks
+
+The active model set has checked-in results for UniMER-Test, MathWriting test, and OpenStax mixed pages. See the [published benchmark summary](https://github.com/SakuraMathcraft/MathCraft-Models/tree/main/benchmarks) and the [full reproducibility suite](https://github.com/SakuraMathcraft/LaTeXSnipper/tree/main/benchmarks/mathcraft_ocr).
+
 ## Development
 
 Run tests from the repository root:

{mathcraft_ocr-0.2.3 → mathcraft_ocr-0.2.5}/README.md

@@ -6,7 +6,7 @@ MathCraft OCR recognizes formulae, text, and mixed mathematical documents with a
 
 ## Quick Start
 
-Current PyPI release line: `mathcraft-ocr 0.2.3`.
+Current PyPI release line: `mathcraft-ocr 0.2.5`.
 
 Install the library and CLI without choosing an ONNX Runtime backend:
 
@@ -121,11 +121,13 @@ for block in result.blocks:
 
 ## Runtime Release Notes
 
-`mathcraft-ocr 0.2.3` fixes cross-platform hardware sizing. Memory detection now uses optional `psutil`, Windows API, POSIX `sysconf`, and macOS `vm_stat`, so CPU batch sizing no longer falls back to Windows-only memory data on Linux or macOS. It also moves the default writable model cache to platform-native user data locations on macOS and Linux. The active `v1.0.0` ONNX model asset set is unchanged.
+`mathcraft-ocr 0.2.5` improves multiline formula recognition by retaining aligned short continuation rows, excluding dark screenshot frames, and comparing uncertain wide-line segments with whole-line recognition. The `v1.0.0` formula-recognition asset now uses the MathCraft-owned `mathcraft-formula-rec` identity in its configuration; the ONNX graphs and weights are unchanged.
 
-Earlier `0.2.x` releases improved runtime-side formula post-processing without changing the active `v1.0.0` ONNX model asset set. They keep compact fraction expressions whole, avoid splitting matrix-like wide formulas, add relation-aware `aligned` output, and retry severe segmented-line artifacts with safer whole-line or whole-image recognition.
+`mathcraft-ocr 0.2.4` fixes DirectML provider handling for RapidOCR text recognition. CUDA and TensorRT providers enable CUDA runtime options, DirectML enables DirectML runtime options, and CPU remains CPU-only. The active `v1.0.0` ONNX model asset set is unchanged.
+
+`mathcraft-ocr 0.2.3` fixes cross-platform hardware sizing. Memory detection now uses optional `psutil`, Windows API, POSIX `sysconf`, and macOS `vm_stat`, so CPU batch sizing no longer falls back to Windows-only memory data on Linux or macOS. It also moves the default writable model cache to platform-native user data locations on macOS and Linux.
 
-Earlier `0.2.x` releases updated the runtime-side formula recognition path with line-aware formula splitting and reassembly. This improved long multi-line formula handling without changing the active `v1.0.0` ONNX model asset set.
+Earlier `0.2.x` releases improved runtime-side formula post-processing without changing the active `v1.0.0` ONNX model asset set. They keep compact fraction expressions whole, avoid splitting matrix-like wide formulas, add relation-aware `aligned` output, and retry severe segmented-line artifacts with safer whole-line or whole-image recognition.
 
 ## Model Set
 
@@ -145,9 +147,14 @@ mathcraft-formula-det.zip
 mathcraft-formula-rec.zip
 mathcraft-text-det.zip
 mathcraft-text-rec.zip
+models.v1.json
 SHA256SUMS.txt
 ```
 
+The runtime downloads the four model archives. `models.v1.json` records their
+contents and source URLs, while `SHA256SUMS.txt` provides archive-level integrity
+checks for release verification.
+
 Default writable model root:
 
 ```text
@@ -194,27 +201,19 @@ Sparse title/cover-style page used to check layout stability.
 
 <img width="1221" height="1898" alt="debug_blocks" src="https://github.com/user-attachments/assets/6c6404e0-bea4-4811-b135-feff3a063a18" />
 
-## Benchmark Snapshot
+## Reproducible Benchmarks
 
-Local `block_layout_regression_v4` telemetry:
+| Benchmark | Scale | Reported result |
+| --- | ---: | --- |
+| UniMER-Test | 23,757 formulas | BLEU-4 `0.7946`; official CDM `0.9288` |
+| MathWriting test | 7,644 samples | BLEU-4 `0.5467`; official CDM `0.750`; render success `98.63%` |
+| OpenStax mixed pages | 200 pages | `0` failures, `0` empty outputs; median `6.65 s/page` |
 
-| Metric | Value |
-| --- | ---: |
-| Pages | 10 |
-| Total blocks | 495 |
-| Text characters | 21,417 |
-| Markdown lines | 304 |
-| Mean page time | 8.34 s |
-| Fastest page | 1.33 s |
-| Slowest page | 18.53 s |
+All recorded runs used `CUDAExecutionProvider`. The datasets cover different tasks and protocols, so the rows are not a model-ranking comparison.
 
-Environment:
+![Published-context UniMER-Test BLEU comparison](benchmarks/figures/unimer_public_bleu_comparison.png)
 
-```text
-Provider: CUDAExecutionProvider
-Runtime: MathCraft OCR v1
-Backend: ONNX Runtime
-```
+See the [benchmark report, charts, provenance, and reproduction notes](benchmarks/README.md).
 
 ## Why It Is Stable
 

{mathcraft_ocr-0.2.3 → mathcraft_ocr-0.2.5}/README_MATHCRAFT_OCR.md

@@ -150,6 +150,10 @@ Model artifacts are downloaded from the MathCraft-Models release assets declared
 
 The actual provider is available on results through the `provider` field.
 
+## Benchmarks
+
+The active model set has checked-in results for UniMER-Test, MathWriting test, and OpenStax mixed pages. See the [published benchmark summary](https://github.com/SakuraMathcraft/MathCraft-Models/tree/main/benchmarks) and the [full reproducibility suite](https://github.com/SakuraMathcraft/LaTeXSnipper/tree/main/benchmarks/mathcraft_ocr).
+
 ## Development
 
 Run tests from the repository root:

{mathcraft_ocr-0.2.3 → mathcraft_ocr-0.2.5}/mathcraft_ocr/__init__.py

@@ -2,7 +2,7 @@
 
 from __future__ import annotations
 
-__version__ = "0.2.3"
+__version__ = "0.2.5"
 
 __all__ = [
     "DoctorReport",

{mathcraft_ocr-0.2.3 → mathcraft_ocr-0.2.5}/mathcraft_ocr/adapters/text_recognizer.py

@@ -54,12 +54,14 @@ def recognize_pp_text_lines(
 
 def _create_pp_text_recognizer(model_dir: Path, provider_info) -> TextRecognizer:
     model_dir = model_dir.resolve()
-    use_cuda = bool(getattr(provider_info, "device", "") == "gpu")
-    return _create_pp_text_recognizer_cached(str(model_dir), use_cuda)
+    active_provider = str(getattr(provider_info, "active_provider", "") or "")
+    use_cuda = active_provider in {"CUDAExecutionProvider", "TensorrtExecutionProvider"}
+    use_dml = active_provider == "DmlExecutionProvider"
+    return _create_pp_text_recognizer_cached(str(model_dir), use_cuda, use_dml)
 
 
 @lru_cache(maxsize=8)
-def _create_pp_text_recognizer_cached(model_dir: str, use_cuda: bool) -> TextRecognizer:
+def _create_pp_text_recognizer_cached(model_dir: str, use_cuda: bool, use_dml: bool) -> TextRecognizer:
     model_dir = Path(model_dir)
     model_candidates = sorted(model_dir.glob("**/*rec*.onnx"))
     if not model_candidates:
@@ -96,7 +98,7 @@ def _create_pp_text_recognizer_cached(model_dir: str, use_cuda: bool) -> TextRec
                 "cudnn_conv_algo_search": "EXHAUSTIVE",
                 "do_copy_in_default_stream": True,
             },
-            "use_dml": False,
+            "use_dml": use_dml,
             "dm_ep_cfg": None,
             "use_cann": False,
             "cann_ep_cfg": {

{mathcraft_ocr-0.2.3 → mathcraft_ocr-0.2.5}/mathcraft_ocr/formula_lines.py

@@ -2,6 +2,7 @@
 
 from __future__ import annotations
 
+from collections import deque
 from dataclasses import dataclass
 import re
 
@@ -113,6 +114,7 @@ def _ink_mask(rgb: np.ndarray) -> np.ndarray:
     background = float(np.percentile(gray, 95))
     threshold = min(245.0, max(80.0, background - 28.0))
     mask = gray < threshold
+    mask = _remove_dark_outer_frame(mask)
 
     height, width = mask.shape
     border = max(1, min(height, width) // 80)
@@ -124,6 +126,50 @@ def _ink_mask(rgb: np.ndarray) -> np.ndarray:
     return mask
 
 
+def _remove_dark_outer_frame(mask: np.ndarray) -> np.ndarray:
+    height, width = mask.shape
+    corner_size = max(4, min(height, width) // 30)
+    corners = (
+        mask[:corner_size, :corner_size],
+        mask[:corner_size, -corner_size:],
+        mask[-corner_size:, :corner_size],
+        mask[-corner_size:, -corner_size:],
+    )
+    dark_corners = sum(float(corner.mean()) >= 0.25 for corner in corners)
+    if dark_corners < 3:
+        return mask
+
+    connected = np.zeros_like(mask, dtype=bool)
+    pending: deque[tuple[int, int]] = deque()
+    for y_start, x_start in (
+        (0, 0),
+        (0, width - corner_size),
+        (height - corner_size, 0),
+        (height - corner_size, width - corner_size),
+    ):
+        points = np.argwhere(mask[y_start : y_start + corner_size, x_start : x_start + corner_size])
+        for y_offset, x_offset in points:
+            y = y_start + int(y_offset)
+            x = x_start + int(x_offset)
+            if connected[y, x]:
+                continue
+            connected[y, x] = True
+            pending.append((y, x))
+
+    while pending:
+        y, x = pending.popleft()
+        for next_y in range(max(0, y - 1), min(height, y + 2)):
+            for next_x in range(max(0, x - 1), min(width, x + 2)):
+                if connected[next_y, next_x] or not mask[next_y, next_x]:
+                    continue
+                connected[next_y, next_x] = True
+                pending.append((next_y, next_x))
+
+    if int(connected.sum()) < max(32, int(round(mask.size * 0.0005))):
+        return mask
+    return mask & ~connected
+
+
 def _row_bands(row_has_ink: np.ndarray) -> list[tuple[int, int]]:
     bands: list[tuple[int, int]] = []
     start: int | None = None
@@ -169,16 +215,14 @@ def _band_looks_like_formula_row(
         return False
 
     active_columns = int(np.count_nonzero(band.any(axis=0)))
-    if active_columns < max(28, int(round(image_width * 0.13))):
+    if active_columns < max(12, int(round(image_width * 0.015))):
         return False
 
     points = np.argwhere(band)
     if points.size == 0:
         return False
     x_span = int(points[:, 1].max() - points[:, 1].min() + 1)
-    if x_span >= int(round(image_width * 0.55)):
-        return active_columns >= max(28, int(round(image_width * 0.08)))
-    return x_span >= int(round(image_width * 0.22))
+    return x_span >= 24
 
 
 def _filter_annotation_bands(
@@ -189,10 +233,44 @@ def _filter_annotation_bands(
 ) -> list[tuple[int, int]]:
     if len(bands) < 2:
         return bands
-    stats = [(_band_active_columns(mask, top, bottom), top, bottom) for top, bottom in bands]
-    max_active = max(active for active, _top, _bottom in stats)
+    stats = []
+    for top, bottom in bands:
+        points = np.argwhere(mask[top : bottom + 1, :])
+        if points.size == 0:
+            continue
+        stats.append(
+            (
+                _band_active_columns(mask, top, bottom),
+                int(points[:, 1].min()),
+                bottom - top + 1,
+                top,
+                bottom,
+            )
+        )
+    if len(stats) < 2:
+        return [(top, bottom) for _active, _left, _height, top, bottom in stats]
+
+    max_active = max(active for active, _left, _height, _top, _bottom in stats)
     min_active = max(int(round(image_width * 0.13)), int(round(max_active * 0.35)))
-    return [(top, bottom) for active, top, bottom in stats if active >= min_active]
+    strong = [stat for stat in stats if stat[0] >= min_active]
+    if not strong:
+        return []
+
+    strong_heights = sorted(height for _active, _left, height, _top, _bottom in strong)
+    median_height = strong_heights[len(strong_heights) // 2]
+    aligned_tolerance = max(10, int(round(image_width * 0.03)))
+    minimum_context_height = max(5, int(round(median_height * 0.40)))
+    strong_lefts = [left for _active, left, _height, _top, _bottom in strong]
+
+    kept: list[tuple[int, int]] = []
+    for active, left, height, top, bottom in stats:
+        if active >= min_active:
+            kept.append((top, bottom))
+            continue
+        left_aligned = any(abs(left - strong_left) <= aligned_tolerance for strong_left in strong_lefts)
+        if height >= minimum_context_height and left_aligned:
+            kept.append((top, bottom))
+    return kept
 
 
 def _looks_like_compact_fraction_split(
@@ -226,7 +304,15 @@ def _looks_like_compact_fraction_split(
         first_width <= image_width * 0.86 and first_inset >= image_width * 0.06
     )
     second_is_wider = second_width >= first_width * 1.15 and second_inset <= image_width * 0.08
-    return first_is_centered_fragment and second_is_wider
+    centers_are_close = abs(
+        (first_left + first_right) - (second_left + second_right)
+    ) <= image_width * 0.12
+    widths_are_comparable = (
+        second_width >= first_width * 0.90
+        and second_inset >= image_width * 0.04
+        and centers_are_close
+    )
+    return first_is_centered_fragment and (second_is_wider or widths_are_comparable)
 
 
 def _band_bounds(mask: np.ndarray, top: int, bottom: int) -> tuple[int, int, int, int] | None:

{mathcraft_ocr-0.2.3 → mathcraft_ocr-0.2.5}/mathcraft_ocr/manifests/models.v1.json

@@ -20,7 +20,7 @@
       "files": [
         {
           "path": "config.json",
-          "sha256": "9f3812441d397c871b9b2a74e8d956b939aec5f4f45745bba9214e968d56449d"
+          "sha256": "cc60b5e3fa221c49086147e82583b3bf62a62917bddf38decfccf7ccfdb1e53b"
         },
         {
           "path": "encoder_model.onnx",

{mathcraft_ocr-0.2.3 → mathcraft_ocr-0.2.5}/mathcraft_ocr/providers.py

@@ -25,6 +25,14 @@ class ProviderInfo:
     gpu_runtime_ok: bool
     cpu_fallback: bool
 
+    @property
+    def use_cuda(self) -> bool:
+        return self.active_provider in {"CUDAExecutionProvider", "TensorrtExecutionProvider"}
+
+    @property
+    def use_dml(self) -> bool:
+        return self.active_provider == "DmlExecutionProvider"
+
 
 def detect_providers(prefer: str = "auto") -> ProviderInfo:
     prefer_norm = (prefer or "auto").strip().lower()

{mathcraft_ocr-0.2.3 → mathcraft_ocr-0.2.5}/mathcraft_ocr/runtime.py

@@ -81,6 +81,8 @@ ONNX_WARMUP_HANDLERS = {
 }
 
 FORMULA_MAX_NEW_TOKENS = 512
+_LOW_SEGMENT_CONFIDENCE = 0.95
+_WHOLE_LINE_CONFIDENCE_MARGIN = 0.02
 
 
 class MathCraftRuntime:
@@ -482,7 +484,7 @@ class MathCraftRuntime:
                 count = len(line_group.crops)
                 grouped_results.append(flat_results[offset : offset + count])
                 offset += count
-            grouped_results = _repair_severe_segmented_lines(
+            grouped_results = _repair_uncertain_segmented_lines(
                 rgb,
                 line_groups,
                 grouped_results,
@@ -621,7 +623,7 @@ def _merge_formula_group_results(results: list[list[tuple[str, float]]]) -> tupl
     return text, score
 
 
-def _repair_severe_segmented_lines(
+def _repair_uncertain_segmented_lines(
     rgb,
     line_groups,
     grouped_results: list[list[tuple[str, float]]],
@@ -635,7 +637,10 @@ def _repair_severe_segmented_lines(
         if len(line_group.crops) <= 1:
             continue
         line_text = compose_formula_line([text for text, _score in line_results])
-        if not has_severe_latex_quality_issue(line_text):
+        has_low_confidence = any(
+            score < _LOW_SEGMENT_CONFIDENCE for _text, score in line_results
+        )
+        if not has_low_confidence and not has_severe_latex_quality_issue(line_text):
             continue
         fallback_text, fallback_score = recognize_formula_image(
             _crop_formula_line_group(rgb, line_group),
@@ -644,7 +649,18 @@ def _repair_severe_segmented_lines(
             max_new_tokens=max_new_tokens,
         )
         line_score = _mean_score(line_results)
-        if _prefer_formula_fallback(line_text, line_score, fallback_text, fallback_score):
+        fallback_is_cleaner = _prefer_formula_fallback(
+            line_text,
+            line_score,
+            fallback_text,
+            fallback_score,
+        )
+        fallback_is_more_confident = (
+            has_low_confidence
+            and not has_severe_latex_quality_issue(fallback_text)
+            and fallback_score > line_score + _WHOLE_LINE_CONFIDENCE_MARGIN
+        )
+        if fallback_is_cleaner or fallback_is_more_confident:
             repaired[index] = [(fallback_text, fallback_score)]
     return repaired
 

{mathcraft_ocr-0.2.3 → mathcraft_ocr-0.2.5}/mathcraft_ocr/serialization.py

@@ -72,6 +72,8 @@ def provider_info_to_json(provider_info) -> dict:
         "gpu_requested": provider_info.gpu_requested,
         "gpu_runtime_ok": provider_info.gpu_runtime_ok,
         "cpu_fallback": provider_info.cpu_fallback,
+        "use_cuda": provider_info.use_cuda,
+        "use_dml": provider_info.use_dml,
     }
 
 

{mathcraft_ocr-0.2.3 → mathcraft_ocr-0.2.5/mathcraft_ocr.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: mathcraft-ocr
-Version: 0.2.3
+Version: 0.2.5
 Summary: ONNX-only OCR runtime for mathematical documents
 Author: SakuraMathcraft
 License-Expression: MIT
@@ -196,6 +196,10 @@ Model artifacts are downloaded from the MathCraft-Models release assets declared
 
 The actual provider is available on results through the `provider` field.
 
+## Benchmarks
+
+The active model set has checked-in results for UniMER-Test, MathWriting test, and OpenStax mixed pages. See the [published benchmark summary](https://github.com/SakuraMathcraft/MathCraft-Models/tree/main/benchmarks) and the [full reproducibility suite](https://github.com/SakuraMathcraft/LaTeXSnipper/tree/main/benchmarks/mathcraft_ocr).
+
 ## Development
 
 Run tests from the repository root:

{mathcraft_ocr-0.2.3 → mathcraft_ocr-0.2.5}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "mathcraft-ocr"
-version = "0.2.3"
+version = "0.2.5"
 description = "ONNX-only OCR runtime for mathematical documents"
 readme = "README_MATHCRAFT_OCR.md"
 requires-python = ">=3.10"