From 9434f2b22bbc3a397f881fa812a98931f6d67e9f Mon Sep 17 00:00:00 2001 From: wren <“porlong@qq.com”> Date: Fri, 22 May 2026 14:41:42 +0800 Subject: [PATCH] fix: improve page quality vlm detection --- .../leaudit_bridge/resilient_clients.py | 2 +- .../services/impl/pageQualityServiceImpl.py | 81 ++++++++++++++++-- tests/test_page_quality_vlm.py | 83 +++++++++++++++++++ 3 files changed, 159 insertions(+), 7 deletions(-) diff --git a/fastapi_modules/fastapi_leaudit/leaudit_bridge/resilient_clients.py b/fastapi_modules/fastapi_leaudit/leaudit_bridge/resilient_clients.py index 369b271..fb42ab7 100644 --- a/fastapi_modules/fastapi_leaudit/leaudit_bridge/resilient_clients.py +++ b/fastapi_modules/fastapi_leaudit/leaudit_bridge/resilient_clients.py @@ -239,7 +239,7 @@ class ResilientQwenVLMClient(QwenVLMClient): body = response.json() text = (body.get("choices") or [{}])[0].get("message", {}).get("content", "") parsed = _parse_json_loose(text) - return parsed if isinstance(parsed, dict) else {} + return parsed if isinstance(parsed, dict) else {"result": text, "reason": text} class ResilientChandraOCRClient(ChandraOCRClient): diff --git a/fastapi_modules/fastapi_leaudit/services/impl/pageQualityServiceImpl.py b/fastapi_modules/fastapi_leaudit/services/impl/pageQualityServiceImpl.py index 77e7941..3fe1f78 100644 --- a/fastapi_modules/fastapi_leaudit/services/impl/pageQualityServiceImpl.py +++ b/fastapi_modules/fastapi_leaudit/services/impl/pageQualityServiceImpl.py @@ -6,6 +6,7 @@ from typing import Any from pathlib import Path import tempfile import logging +import json import fitz from leaudit.converters import doc2pdf @@ -30,9 +31,11 @@ _PAGE_QUALITY_VLM_PROMPT = """ 你是文档扫描图片质量检测员。请判断这 1 页文档图片是否适合继续做 OCR 与合同/公文评查。 判定标准： -1. pass：文字主体清晰、方向正常、没有明显截断，能稳定阅读。 -2. review：存在轻微模糊、倾斜、阴影、低对比度、局部遮挡、轻微截断，建议人工确认但仍可能可读。 -3. reject：严重模糊、重影、过曝/过暗、页面大面积缺失、关键文字不可辨认、方向严重错误、空白页或非文档页，建议重拍。 +1. 必须同时检查整页扫描质量，以及页面内所有内嵌照片、证据照片、现场照片、截图、印章和签名图片的清晰度。 +2. pass：文字主体清晰、方向正常、没有明显截断；页面内嵌照片/证据照片也能辨认关键视觉信息。 +3. review：存在轻微模糊、倾斜、阴影、低对比度、局部遮挡、轻微截断；或内嵌照片/证据照片主体明显发虚、牌匾/场所/人物/关键物证不易辨认，建议人工确认但仍可能可用。 +4. reject：严重模糊、重影、过曝/过暗、页面大面积缺失、关键文字不可辨认、方向严重错误、空白页或非文档页；或内嵌证据照片主体无法辨认、关键证据信息不可用，建议重拍。 +5. 即使页面周边文字清楚，只要内嵌证据照片明显模糊，也不能判 pass，至少判 review，严重时判 reject。 只输出 JSON，不要输出 Markdown，不要解释额外文本： {"status":"pass|review|reject","score":0.0到1.0,"reason":"20字以内中文原因"} @@ -495,12 +498,28 @@ class PageQualityServiceImpl(IPageQualityService): logger.warning("VLM page quality detection failed: %s", exc) return "review", 0.5, "VLM图片质量检测失败，需人工确认" - status = str((result or {}).get("status") or "").strip().lower() + result_dict = self._coerce_vlm_result(result) + status = self._normalize_quality_status( + self._first_non_empty( + result_dict, + ("status", "quality_status", "qualityStatus", "result", "label", "decision", "conclusion"), + ) + ) + reason = self._normalize_quality_reason( + self._first_non_empty( + result_dict, + ("reason", "quality_reason", "qualityReason", "message", "msg", "detail", "explanation", "description"), + ) + ) + if status is None and reason: + status = self._normalize_quality_status(reason) if status not in {"pass", "review", "reject"}: return "review", 0.5, "VLM返回结果不可用，需人工确认" - score = self._normalize_quality_score((result or {}).get("score"), status) - reason = str((result or {}).get("reason") or "").strip() or None + score = self._normalize_quality_score( + self._first_non_empty(result_dict, ("score", "quality_score", "qualityScore", "confidence")), + status, + ) if status != "pass" and not reason: reason = "页面图片质量需人工确认" return status, score, reason @@ -526,6 +545,56 @@ class PageQualityServiceImpl(IPageQualityService): return defaults[status] return max(0.0, min(1.0, score)) + def _coerce_vlm_result(self, result: Any) -> dict[str, Any]: + if isinstance(result, dict): + return result + if isinstance(result, str): + text_result = result.strip() + if not text_result: + return {} + try: + parsed = json.loads(text_result) + except json.JSONDecodeError: + return {"result": text_result, "reason": text_result} + return parsed if isinstance(parsed, dict) else {"result": text_result} + return {} + + def _first_non_empty(self, payload: dict[str, Any], keys: tuple[str, ...]) -> Any: + for key in keys: + value = payload.get(key) + if value is not None and str(value).strip(): + return value + return None + + def _normalize_quality_status(self, raw_status: Any) -> str | None: + text_status = str(raw_status or "").strip().lower() + if not text_status: + return None + compact_status = text_status.replace(" ", "").replace("_", "").replace("-", "") + if compact_status in {"pass", "passed", "ok", "good", "clear", "readable"}: + return "pass" + if compact_status in {"review", "warn", "warning", "manual", "uncertain", "suspect", "suspicious"}: + return "review" + if compact_status in {"reject", "rejected", "fail", "failed", "bad", "unreadable", "retake"}: + return "reject" + + reject_keywords = ("不通过", "拒绝", "重拍", "不可读", "无法辨认", "无法识别", "严重", "大面积缺失", "空白页") + review_keywords = ("复核", "人工", "疑似", "轻微", "建议确认", "建议人工", "模糊", "不清晰", "低对比", "发虚") + pass_keywords = ("通过", "合格", "清晰", "可读") + if any(keyword in text_status for keyword in reject_keywords): + return "reject" + if any(keyword in text_status for keyword in review_keywords): + return "review" + if any(keyword in text_status for keyword in pass_keywords): + return "pass" + return None + + def _normalize_quality_reason(self, raw_reason: Any) -> str | None: + reason = str(raw_reason or "").strip() + if not reason: + return None + return reason[:80] + def _document_service(self): if self.DocumentService is None: from fastapi_modules.fastapi_leaudit.services.impl.documentServiceImpl import DocumentServiceImpl diff --git a/tests/test_page_quality_vlm.py b/tests/test_page_quality_vlm.py index 7e8db82..721e090 100644 --- a/tests/test_page_quality_vlm.py +++ b/tests/test_page_quality_vlm.py @@ -1,5 +1,7 @@ import pytest +import httpx +from fastapi_modules.fastapi_leaudit.leaudit_bridge.resilient_clients import ResilientQwenVLMClient from fastapi_modules.fastapi_leaudit.services.impl.pageQualityServiceImpl import PageQualityServiceImpl @@ -32,6 +34,58 @@ async def test_vlm_page_quality_reject_result_is_used(): assert score == 0.18 assert "严重模糊" in reason assert "只输出 JSON" in service.VlmClient.prompts[0][0] + assert "内嵌照片" in service.VlmClient.prompts[0][0] + assert "即使页面周边文字清楚" in service.VlmClient.prompts[0][0] + + +@pytest.mark.asyncio +async def test_vlm_page_quality_embedded_evidence_blur_cannot_pass(): + service = PageQualityServiceImpl() + service.VlmClient = _FakeVlmClient( + { + "quality_status": "疑似模糊", + "quality_score": "0.42", + "message": "内嵌证据照片主体发虚，门头文字不易辨认", + } + ) + + status, score, reason = await service._classify_page_image_by_vlm(b"image-bytes") + + assert status == "review" + assert score == 0.42 + assert "内嵌证据照片" in reason + + +@pytest.mark.asyncio +async def test_vlm_page_quality_chinese_reject_status_is_supported(): + service = PageQualityServiceImpl() + service.VlmClient = _FakeVlmClient( + { + "result": "不通过", + "confidence": 0.1, + "detail": "证据照片严重模糊，关键场所无法辨认", + } + ) + + status, score, reason = await service._classify_page_image_by_vlm(b"image-bytes") + + assert status == "reject" + assert score == 0.1 + assert "严重模糊" in reason + + +@pytest.mark.asyncio +async def test_vlm_page_quality_json_string_result_is_supported(): + service = PageQualityServiceImpl() + service.VlmClient = _FakeVlmClient( + '{"status":"review","score":0.33,"reason":"页面内照片模糊"}' + ) + + status, score, reason = await service._classify_page_image_by_vlm(b"image-bytes") + + assert status == "review" + assert score == 0.33 + assert reason == "页面内照片模糊" @pytest.mark.asyncio @@ -56,3 +110,32 @@ async def test_vlm_page_quality_error_falls_back_to_review_not_pass(): assert status == "review" assert score == 0.5 assert "VLM图片质量检测失败" in reason + + +@pytest.mark.asyncio +async def test_resilient_vlm_extract_multifield_keeps_raw_text_when_json_parse_fails(monkeypatch): + client = ResilientQwenVLMClient(base_url="http://example.test", api_key="x", model="vlm-test") + + async def fake_post_with_retry(payload): + return httpx.Response( + 200, + json={ + "choices": [ + { + "message": { + "content": "疑似模糊：内嵌证据照片主体发虚，建议人工复核", + } + } + ] + }, + ) + + monkeypatch.setattr(client, "_post_with_retry", fake_post_with_retry) + + result = await client.extract_multifield( + prompt="图片质量检测", + images_data_urls=["data:image/png;base64,xxx"], + ) + + assert result["result"].startswith("疑似模糊") + assert "内嵌证据照片" in result["reason"]