增加PDF采样率配置参数 pdf转图片dpi为72*scale

src/bisheng_unstructured/api/pipeline.py

@@ -24,7 +24,7 @@
 from bisheng_unstructured.staging.base import convert_to_isd
 
 
-def partition_pdf(filename, model_params, **kwargs):
+def partition_pdf(filename, model_params,scale, **kwargs):
     if kwargs.get("mode") == "local":
         # pypdf 进行解析
         import pypdf
@@ -43,7 +43,7 @@ def partition_pdf(filename, model_params, **kwargs):
         # if rt_type in {"ocr_sdk", "idp"}:
         #     doc = IDP_PDFDocument(file=filename, model_params=model_params, **kwargs)
         # else:
-        doc = PDFDocument(file=filename, model_params=model_params, **kwargs)
+        doc = PDFDocument(file=filename, model_params=model_params,scale=scale, **kwargs)
 
         _ = doc.pages
         return doc.elements
@@ -156,7 +156,7 @@ def predict(self, inp: UnstructuredInput) -> UnstructuredOutput:
             return UnstructuredOutput(status_code=400, status_message="本地模型不支持图片")
 
         if part_func == partition_pdf or part_func == partition_image:
-            part_inp.update({"model_params": self.pdf_model_params})
+            part_inp.update({"model_params": self.pdf_model_params,"scale":inp.scale})
         try:
             elements = part_func(**part_inp)
             mode = inp.mode

src/bisheng_unstructured/api/types.py

@@ -11,6 +11,7 @@ class UnstructuredInput(BaseModel):
     mode: str = "text"  # text, partition, vis, topdf
     file_path: Optional[str] = None
     file_type: Optional[str] = None
+    scale: Optional[int] = 1
     is_scan: Optional[bool] = None
 
 

src/bisheng_unstructured/config/config.yaml

@@ -11,13 +11,13 @@ logger_conf:
       # 日志级别
       level: INFO
       # 和原生不一样，后端会将配置使用eval()执行转为函数用来过滤特定日志级别。推荐lambda
-      filter: "lambda record: record['level'].name == 'INFO'"
+      # filter: "lambda record: record['level'].name == 'INFO'"
       # 日志格式化函数，extra内支持trace_id
       format: "[{time:YYYY-MM-DD HH:mm:ss.SSSSSS}]|{level}|BISHENG|{extra[trace_id]}|{process.id}|{thread.id}|{message}"
       # 每天的几点进行切割
       rotation: "00:00"
       retention: "3 Days"
-    - sink: "/app/logs/err-v0-BISHENG-{HOSTNAME}.log"
+    - sink: "/app/logs/err-v0-BISHENG-UNS-{HOSTNAME}.log"
       level: ERROR
       filter: "lambda record: record['level'].name == 'ERROR'"
       format: "[{time:YYYY-MM-DD HH:mm:ss.SSSSSS}]|{level}|BISHENG|{extra[trace_id]}||{process.id}|{thread.id}|||#EX_ERR:POS={name},line {line},ERR=500,EMSG={message}"
@@ -26,11 +26,11 @@ logger_conf:
 
 # pdf解析需要用到的模型配置, 配置了rt_server环境变量的话会替换为对应的地址
 pdf_model_params:
-  layout_ep: "http://192.168.106.12:9001/v2.1/models/elem_layout_v1/infer"
-  cell_model_ep: "http://192.168.106.12:9001/v2.1/models/elem_table_cell_detect_v1/infer"
-  rowcol_model_ep: "http://192.168.106.12:9001/v2.1/models/elem_table_rowcol_detect_v1/infer"
-  table_model_ep: "http://192.168.106.12:9001/v2.1/models/elem_table_detect_v1/infer"
-  ocr_model_ep: "http://192.168.106.12:9001/v2.1/models/elem_ocr_collection_v3/infer"
+  layout_ep: "http://10.60.38.67:3011/v2.1/models/elem_layout_v1/infer"
+  cell_model_ep: "http://10.60.38.67:3011/v2.1/models/elem_table_cell_detect_v1/infer"
+  rowcol_model_ep: "http://10.60.38.67:3011/v2.1/models/elem_table_rowcol_detect_v1/infer"
+  table_model_ep: "http://10.60.38.67:3011/v2.1/models/elem_table_detect_v1/infer"
+  ocr_model_ep: "http://10.60.38.67:3011/v2.1/models/elem_ocr_collection_v3/infer"
 
 # 是否全部走ocr识别, false的话则由代码逻辑判断是否需要走ocr识别
 is_all_ocr: false

src/bisheng_unstructured/documents/pdf_parser/pdf.py

@@ -271,6 +271,7 @@ def __init__(
         support_formula: bool = False,
         enable_isolated_formula: bool = False,
         n_parallel: int = 10,
+        scale: float = 1,
         **kwargs,
     ) -> None:
         """Initialize with a file path."""
@@ -301,6 +302,7 @@ def __init__(
         self.support_formula = support_formula
         self.enable_isolated_formula = enable_isolated_formula
         self.n_parallel = n_parallel
+        self.scale = scale
         self.is_scan = is_scan
         self.mode = kwargs.get("mode", "local")
         super().__init__()
@@ -321,7 +323,7 @@ def _get_image_blobs(self, fitz_doc, pdf_reader, n=None, start=0):
             except Exception:
                 # some pdf input cannot get render image from fitz
                 page = pdf_reader.get_page(pg)
-                pil_image = page.render().to_pil()
+                pil_image = page.render(scale=self.scale).to_pil()
                 img_byte_arr = io.BytesIO()
                 pil_image.save(img_byte_arr, format="PNG")
                 bytes_img = img_byte_arr.getvalue()
@@ -1102,7 +1104,14 @@ def _get_elem(blocks, is_first=True):
             b0, b0_label, r0, r0_w, r0_h = _get_elem(groups[i], False)
 
         return groups
-
+    def _divide_by_scale(self,input_list,scale):
+        result = []
+        for item in input_list:
+            if isinstance(item, list):
+                result.append(self._divide_by_scale(item,scale))
+            else:
+                result.append(item / scale)
+        return result
     def _save_to_pages(self, groups, page_inds, lang):
         TITLE_ID = 3
         TEXT_ID = 4
@@ -1118,7 +1127,7 @@ def _save_to_pages(self, groups, page_inds, lang):
                 text = b.block_text
 
                 element = None
-                extra_data = {"bboxes": [bbox], "pages": b.pages}
+                extra_data = {"bboxes": [self._divide_by_scale(bbox,self.scale)], "pages": b.pages}
 
                 if label == TABLE_ID:
                     # html = b[-1]
@@ -1158,7 +1167,7 @@ def _save_to_pages(self, groups, page_inds, lang):
                     # lines = b[-1]
                     lines = b.ts
                     line_cnt = len(lines)
-                    extra_data.update({"bboxes": line_bboxes})
+                    extra_data.update({"bboxes": self._divide_by_scale(line_bboxes,self.scale)})
                     if True or lang == "zh":  # for join test only
                         line_chars_cnt = [len(line) for line in lines]
                         indexes = []
@@ -1225,7 +1234,7 @@ def _task(textpage_info, bytes_img, img, is_scan, lang, rot_matirx, page_index:
             page_imgs = []
             for idx in range(start, start + n):
                 page = pdf_doc.get_page(idx)
-                pil_image = page.render().to_pil()
+                pil_image = page.render(scale=self.scale).to_pil()
                 page_imgs.append(pil_image)
                 img_byte_arr = io.BytesIO()
                 pil_image.save(img_byte_arr, format="PNG")
@@ -1261,7 +1270,7 @@ def _task(textpage_info, bytes_img, img, is_scan, lang, rot_matirx, page_index:
                     type_texts = "".join(type_texts)
                     zh_n = len(re.findall(ZH_CHAR, type_texts))
                     total_n = len(type_texts)
-                    is_scan = total_n < 200
+                    is_scan = total_n < 200 or self.scale!=1
                     if not is_scan:
                         lang = "zh" if zh_n > 200 or zh_n / total_n > 0.5 else "eng"
                     else:

src/bisheng_unstructured/models/ocr_agent.py

@@ -29,21 +29,21 @@
     },
     "scene_mapping": {
         "print": {
-            "det": "general_text_det_v2.0",
-            "recog": "general_text_reg_nb_v1.0_faster",
+            "det": "general_text_det_mrcnn_v2.0",
+            "recog": "transformer-blank-v0.2-faster",
         },
         "hand": {
-            "det": "general_text_det_v2.0",
-            "recog": "general_text_reg_nb_v1.0_faster",
+            "det": "general_text_det_mrcnn_v2.0",
+            "recog": "transformer-hand-v1.16-faster",
         },
         "print_recog": {
-            "recog": "general_text_reg_nb_v1.0_faster",
+            "recog": "transformer-blank-v0.2-faster",
         },
         "hand_recog": {
-            "recog": "general_text_reg_nb_v1.0_faster",
+            "recog": "transformer-hand-v1.16-faster",
         },
         "det": {
-            "det": "general_text_det_v2.0",
+            "det": "general_text_det_mrcnn_v2.0",
         },
     },
 }

tests/test_pdf_parser.py

@@ -5,7 +5,7 @@
 from bisheng_unstructured.documents.pdf_parser.pdf import PDFDocument
 
 RT_EP = os.environ.get("RT_EP", "192.168.106.12:9001")
-TEST_RT_URL = f"http://{RT_EP}/v2.1/models/"
+TEST_RT_URL = f"http://10.60.38.67:3011/v2.1/models/"
 
 
 def test_pdf_doc():
@@ -22,8 +22,8 @@ def test_pdf_doc():
         "table_model_ep": table_model_ep,
     }
 
-    filename = "examples/docs/layout-parser-paper-fast.pdf"
-    pdf_doc = PDFDocument(file=filename, model_params=model_params, n=2)
+    filename = "examples/docs/模糊+水印-租赁物清单-20231213142249768106738.pdf"
+    pdf_doc = PDFDocument(file=filename, model_params=model_params, n=2,scale=2)
     pages = pdf_doc.pages
     elements = pdf_doc.elements
 
@@ -156,9 +156,9 @@ def test_pdf_doc7():
         "ocr_model_ep": f"{TEST_RT_URL}elem_ocr_collection_v3/infer",
     }
 
-    filename = "examples/docs/maoxuan_scan.pdf"
+    filename = "examples/docs/模糊+水印-租赁物清单-20231213142249768106738.pdf"
     pdf_doc = PDFDocument(
-        file=filename, model_params=model_params, enhance_table=False, start=0, n=100
+        file=filename, model_params=model_params, enhance_table=False, start=0, n=100,scale=2
     )
     pages = pdf_doc.pages
     elements = pdf_doc.elements
@@ -276,9 +276,9 @@ def test_pdf_doc10():
 # test_pdf_doc5()
 # test_pdf_doc6()
 
-# test_pdf_doc7()
+test_pdf_doc7()
 # test_pdf_doc8()
 # test_pdf_doc9()
 # test_regress()
 
-test_pdf_doc10()
+# test_pdf_doc10()